research papers in forensic biology

• You are here:
• American Chemical Society
• Discover Chemistry
• Science Releases

Recent advances in forensic science research

For immediate release, acs news service weekly presspac: april 20, 2022.

Forensic scientists collect and analyze evidence during a criminal investigation to identify victims, determine the cause of death and figure out “who done it.” Below are some recent papers published in ACS journals reporting on new advances that could help forensic scientists solve crimes. Reporters can request free access to these papers by emailing  newsroom@acs.org .

“Insights into the Differential Preservation of Bone Proteomes in Inhumed and Entombed Cadavers from Italian Forensic Caseworks” Journal of Proteome Research March 22, 2022 Bone proteins can help determine how long ago a person died (post-mortem interval, PMI) and how old they were at the time of their death (age at death, AAD), but the levels of these proteins could vary with burial conditions. By comparing bone proteomes of exhumed individuals who had been entombed in mausoleums or buried in the ground, the researchers found several proteins whose levels were not affected by the burial environment, which they say could help with AAD or PMI estimation.

“Carbon Dot Powders with Cross-Linking-Based Long-Wavelength Emission for Multicolor Imaging of Latent Fingerprints” ACS Applied Nanomaterials Jan. 21, 2022 For decades, criminal investigators have recognized the importance of analyzing latent fingerprints left at crime scenes to help identify a perpetrator, but current methods to make these prints visible have limitations, including low contrast, low sensitivity and high toxicity. These researchers devised a simple way to make fluorescent carbon dot powders that can be applied to latent fingerprints, making them fluoresce under UV light with red, orange and yellow colors.

“Proteomics Offers New Clues for Forensic Investigations” ACS Central Science Oct. 18, 2021 This review article describes how forensic scientists are now turning their attention to proteins in bone, blood or other biological samples, which can sometimes answer questions that DNA can’t. For example, unlike DNA, a person’s complement of proteins (or proteome) changes over time, providing important clues about when a person died and their age at death.

“Integrating the MasSpec Pen with Sub-Atmospheric Pressure Chemical Ionization for Rapid Chemical Analysis and Forensic Applications” Analytical Chemistry May 19, 2021 These researchers previously developed a “MasSpec Pen,” a handheld device integrated with a mass spectrometer for direct analysis and molecular profiling of biological samples. In this article, they develop a new version that can quickly and easily detect and measure compounds, including cocaine, oxycodone and explosives, which can be important in forensics investigations.

The American Chemical Society (ACS) is a nonprofit organization chartered by the U.S. Congress. ACS’ mission is to advance the broader chemistry enterprise and its practitioners for the benefit of Earth and all its people. The Society is a global leader in promoting excellence in science education and providing access to chemistry-related information and research through its multiple research solutions, peer-reviewed journals, scientific conferences, eBooks and weekly news periodical Chemical & Engineering News . ACS journals are among the most cited, most trusted and most read within the scientific literature; however, ACS itself does not conduct chemical research. As a leader in scientific information solutions, its CAS division partners with global innovators to accelerate breakthroughs by curating, connecting and analyzing the world’s scientific knowledge. ACS’ main offices are in Washington, D.C., and Columbus, Ohio.

To automatically receive press releases from the American Chemical Society, contact newsroom@acs.org .

Note: ACS does not conduct research, but publishes and publicizes peer-reviewed scientific studies.

Media Contact

ACS Newsroom newsroom@acs.org

Discover Chemistry  —Menu

• News Releases
• ACS in the News

Accept & Close The ACS takes your privacy seriously as it relates to cookies. We use cookies to remember users, better understand ways to serve them, improve our value proposition, and optimize their experience. Learn more about managing your cookies at Cookies Policy .

1155 Sixteenth Street, NW, Washington, DC 20036, USA |  service@acs.org  | 1-800-333-9511 (US and Canada) | 614-447-3776 (outside North America)

• Terms of Use
• Accessibility

Copyright © 2024 American Chemical Society

Big footsteps and new challenges

• Open access
• Published: 03 May 2022
• Volume 18 , pages 123–124, ( 2022 )

Cite this article

You have full access to this open access article

• Claas T. Buschmann 1 ,
• Biagio Solarino 2 &
• Takahito Hayashi 3  

1797 Accesses

1 Altmetric

Explore all metrics

Avoid common mistakes on your manuscript.

After nearly 15 years, the scientific Chief Editorial Team of Forensic Science, Medicine and Pathology has changed in January 2022 [ 1 , 2 ].

First of all, we gratefully thank former Editor-in-Chief Roger W. Byard, former European Editor Michael Tsokos, and former North American Editor John Hunsaker III for their time and passion invested over the past 15 years in the journal — they have not only read, reviewed, and edited thousands of papers, they have also published extensively themselves and thus contributed to the reputation of the journal, too. Together with the Editorial Board, they made the journal what it is today.

Secondly, we would like to thank the whole team at Springer, who gave us three “newcomers” the opportunity to follow in these big footsteps. We will do our best.

Now brought to the second-most important journal worldwide in forensic medicine with regard to its current impact factor by the former Editorial team, Forensic Science, Medicine and Pathology will continue to explore all aspects of modern-day forensics. The range of topics covered will continue to include international forensic science, medicine, nursing, and pathology, as well as toxicology, human identification, mass disasters/mass war graves, profiling, imaging and forensic radiology, forensic age estimation, policing, wound assessment, child maltreatment, sexual assault, anthropology, archeology, entomology, botany, biology, veterinary pathology, medical-historical forensic research, and DNA. We will continue to insist on high scientific quality of papers in fluent and sufficiently readable English language. We pay great attention to evaluating the mega-authorship reports and ask to limit the number of self-citations, particularly if not necessary.

Mors auxilium vitae (Death Is Help For The Living) , and looking beyond the horizon is crucial in modern forensic medicine. Interdisciplinary questions arising in the daily autopsy routine can be addressed scientifically, and forensic medicine can contribute to walk new paths. Thus, we also welcome “outside-the-box” papers, i.e., scientific research from the interface of forensic medicine and other medical disciplines — there is a significant overlap between forensic medicine and several curative disciplines, especially after a second look [ 3 , 4 , 5 ]. This also applies to the interface of forensic medicine and the judicial system, i.e., legal assessment of forensic findings. As a sub-category of case reports, we have established “From The Court Room” as a brief case description to present and discuss — not necessarily extraordinary — autopsy and/or crime scene features in a specific case and their legal evaluation. What are legal consequences of our work for those affected, and where are the limits of forensic diagnostics? Where can we get better? With this new proposal, we can discuss once again the inference of the forensic publications in a trial [ 6 ]. How does the Judge determine the scientific value of the articles and the qualifications and credentials of a proposed expert witness? Even considering the differences among the legal systems worldwide, we are looking forward to submissions addressing these points.

Forensic Science, Medicine and Pathology will continue to present a balance of forensic research and reviews from around the world to reflect modern advances through peer-reviewed papers, short communications, meeting proceedings, new forensic textbook comments, and case reports. Furthermore, we will open the journal to answers to forensic questions that involve interfaces with other medical disciplines, especially with regard to complications arising from performed — or necessary, but omitted — medical procedures in the broadest sense. Forensic scientists are often involved in medical malpractice lawsuits, healthcare policy, and patient safety management. Therefore, the authors have the opportunity to discuss unusual adverse events, causes of medical malpractice, and the forensic medicine approach to such an interesting field of research. The dead can teach the living.

The scientific future of forensic medicine comprises not only of forensic issues, but involves interdisciplinary cooperation. We intend to be a relevant part of this future — and we can achieve this goal solely with the help of you, the authors and reviewers from all over the world!

Priv.-Doz. Dr. med. Claas T. Buschmann, Kiel/Lübeck, Germany

Editor-in-Chief

Prof. Dr. Biagio Solarino, Bari, Italy

Associate Editor

Prof. Takahito Hayashi, Kagoshima, Japan

Byard RW, Hunsaker JC, Tsokos M. Forensic science medicine and pathology – a change of command. Forensic Sci Med Pathol. 2022;18:116–7.

Article   Google Scholar  

Byard RW. Academic standing and publication. Forensic Sci Med Pathol. 2022;18:1–3.

Buschmann C. More interdisciplinary research is needed in forensic medicine. Forensic Sci Med Pathol. 2019;15:131–2.

Buschmann C, Tsokos M, Kleber C. Preventive pathology: the interface of forensic medicine and trauma surgery for pre-hospital trauma management. Forensic Sci Med Pathol. 2015;11:317–8.

Lacour P, Buschmann C, Storm C, et al. Cardiac implantable electronic device interrogation at forensic autopsy – an underestimated resource? Circulation 2018;137:2730–40.

Jones AW. Highly cited forensic practitioners in the discipline legal and forensic medicine and the importance of peer-review and publication for admission of expert testimony. Forensic Sci Med Pathol. 2022;18:37–44.

Download references

Open Access funding enabled and organized by Projekt DEAL.

Author information

Authors and affiliations.

Institute of Legal Medicine, University Hospital Schleswig-Holstein, Kiel / Lübeck, Germany

Claas T. Buschmann

Institute of Legal Medicine, University of Bari, Bari, Italy

Biagio Solarino

Department of Legal Medicine, Graduate School of Medical and Dental Sciences, Kagoshima University, Kagoshima, Japan

Takahito Hayashi

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Claas T. Buschmann .

Ethics declarations

The editorial was written in compliance with Ethical Standards.

Ethics approval

Ethical approval was not required for this type of paper.

Consent to participate

Informed consent was not required for this type of paper.

Conflict of interest

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Buschmann, C.T., Solarino, B. & Hayashi, T. Big footsteps and new challenges. Forensic Sci Med Pathol 18 , 123–124 (2022). https://doi.org/10.1007/s12024-022-00482-5

Download citation

Accepted : 25 April 2022

Published : 03 May 2022

Issue Date : June 2022

DOI : https://doi.org/10.1007/s12024-022-00482-5

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Find a journal
• Publish with us
• Track your research

The Role of DNA in Forensic Science: A Comprehensive Review

Alketbi, S. K. (2023) The role of DNA in forensic science: A comprehensive review. International Journal of Science and Research Archive, 09 (02), pp. 814-829.

16 Pages Posted: 12 Sep 2023 Last revised: 13 Oct 2023

Salem K Alketbi

Dubai Police - General Department of Forensic Science and Criminology; University of Central Lancashire

Date Written: August 23, 2023

Forensic genetics, leveraging molecular tools and scientific applications, has witnessed significant advancements in DNA analysis over the last three decades. These progressions have enhanced the discrimination power, speed, and sensitivity of DNA profiling methods, enabling the analysis of challenging samples. This article explores the significance of forensic genetics in criminal investigations, traces the historical evolution of DNA analysis techniques, and presents recent developments in the field. The article aims to provide a comprehensive understanding of the crucial role of forensic genetics in criminal investigations and sheds light on the latest trends and breakthroughs in this area. The evolution of DNA typing from ABO blood typing to the current standard of short tandem repeat (STR) analysis is discussed, along with alternative DNA analysis methods, such as Y-chromosome analysis and single nucleotide polymorphism (SNP) typing. Massively parallel sequencing (MPS) represents a groundbreaking advancement, enabling whole genome sequencing and addressing complex cases. The article also covers recent innovations, including DNA methylation analysis, body fluid identification, forensic DNA phenotyping, and genetic genealogy, highlighting their potential benefits in forensic investigations. Despite these advancements, standard STR profiling remains the gold standard due to its established protocols and databases. Ethical considerations regarding data privacy and cost implications are crucial as these technologies continue to progress in their pursuit of justice.

Keywords: Forensic science, Forensic genetics, DNA, DNA Typing, DNA analysis, DNA profiling, Restriction Fragment Length Polymorphism (RFLP), Short Tandem Repeat (STR), STR Typing, Massively Parallel Sequencing (MPS), Next-Generation Sequencing (NGS)

Suggested Citation: Suggested Citation

Salem K Alketbi (Contact Author)

Dubai police - general department of forensic science and criminology ( email ).

Dubai United Arab Emirates

University of Central Lancashire ( email )

Preston United Kingdom

Do you have a job opening that you would like to promote on SSRN?

Paper statistics, related ejournals, biology & law ejournal.

Subscribe to this fee journal for more curated articles on this topic

Biology & Anthropology eJournal

Law enforcement ejournal, forensic science ejournal, criminalistics ejournal.

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

The science and philosophy of the brain and the future of neuroscience.

Conflicts of Interest

• Sepahvand, T.; Power, K.D.; Qin, T.; Yuan, Q. The Basolateral Amygdala: The Core of a Network for Threat Conditioning, Extinction, and Second-Order Threat Conditioning. Biology 2023 , 12 , 1274. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Skolariki, K.; Vrahatis, A.G.; Krokidis, M.G.; Exarchos, T.P.; Vlamos, P. Assessing and Modelling of Post-Traumatic Stress Disorder Using Molecular and Functional Biomarkers. Biology 2023 , 12 , 1050. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Lawson, L.; Spivak, S.; Webber, H.; Yasin, S.; Goncalves, B.; Tarrio, O.; Ash, S.; Ferrol, M.; Ibragimov, A.; Olivares, A.G.; et al. Alterations in Brain Activity Induced by Transcranial Magnetic Stimulation and Their Relation to Decision Making. Biology 2023 , 12 , 1366. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Nguyen, G.H.; Oh, S.; Schneider, C.; Teoh, J.Y.; Engstrom, M.; Santana-Gonzalez, C.; Porter, D.; Quevedo, K. Neurofeedback and Affect Regulation Circuitry in Depressed and Healthy Adolescents. Biology 2023 , 12 , 1399. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Pham, T.Q.; Matsui, T.; Chikazoe, J. Evaluation of the Hierarchical Correspondence between the Human Brain and Artificial Neural Networks: A Review. Biology 2023 , 12 , 1330. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Zhang, R.; Zeng, Y.; Tong, L.; Yan, B. Specific Neural Mechanisms of Self-Cognition and the Application of Brainprint Recognition. Biology 2023 , 12 , 486. [ Google Scholar ] [ CrossRef ] [ PubMed ]

Share and Cite

Keenan, J.P. The Science and Philosophy of the Brain and the Future of Neuroscience. Biology 2024 , 13 , 607. https://doi.org/10.3390/biology13080607

Keenan JP. The Science and Philosophy of the Brain and the Future of Neuroscience. Biology . 2024; 13(8):607. https://doi.org/10.3390/biology13080607

Keenan, Julian Paul. 2024. "The Science and Philosophy of the Brain and the Future of Neuroscience" Biology 13, no. 8: 607. https://doi.org/10.3390/biology13080607

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

The Future of Forensic Biology

• January 2018
• Journal of Biomedicine 3(3):13-18

• Central Forensic Science Laboratory

Abstract and Figures

Discover the world's research

• 25+ million members
• 160+ million publication pages
• 2.3+ billion citations

• Priyanka Kapoor
• Arun Dev Sharma

• Binoy Dolwin

• Sangita Padmakar Rokade
• Geeta Shrivastava

• Shiwani Guleria Sharma
• Sanket Anil Bongale

• Kwang-Mo Choi
• You Hyun Jung
• Eun Kyung Cho

• Rothsovann Yong

• BMC GENOMICS

• Bruce Budowle

• Han Gardeniers
• DEV GENES EVOL

• J FORENSIC SCI

• Patricia Albani

• Recruit researchers
• Join for free
• Login Email Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google Welcome back! Please log in. Email · Hint Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google No account? Sign up

A Forensic Archaeologist Solved the 100-Year-Old Mystery of a Missing WWI Soldier

The family of Private 1st Class Charles McAllister finally has closure.

The level of carnage in World War I led many to believe it was “the War to End All Wars.” But for some people on the home front, there was no certainty—only the question of what had become of their loved ones who had gone off to fight.

For over a century, one family has lived with the mystery of what happened to their soldier who never returned, nor had his death confirmed. Finally, his descendants have received closure as his remains have been identified and brought home—and the forensic expert who played a crucial role in the discovery has told the story of how it all came together.

Forensic archaeologist Jay Silverstein appeared as a guest author for IFLScience to explain how the body of Private 1st Class Charles McAllister, who fought for the U.S. in the Franco-American counter-offensive at Aisne-Marne, was finally identified.

The battle, which occurred on July 18, 1918, resulted in “more than 1,000 U.S. soldiers” unaccounted for, according to Silverstein’s article. “But 85 years later, French archaeologists conducting salvage work ahead of a construction job on what would have been the centre of the battlefield encountered the remains of two American soldiers.”

Those two soldiers’ remains were turned over to the U.S. military’s central identification laboratory (CIL), where Silverstein was working at the time, in 2004. One of the soldiers, Private Francis Lupo, was swiftly identified thanks to his name being embossed on his wallet. In 2006, Lupo would be laid to rest, with full military honors, in Arlington National Cemetery.

The other soldier, then known only as CIL 2004-101-I-02, was deemed impossible to identify at the time, according to Silverstein. “But some 14 years later,” Silverstein writes, “as we approached the 100-year anniversary of the death of this soldier and the end of the first world war, I reopened the case.”

Undertaking this effort on his own time, Silverstein felt that he could use certain factors, including “the date and location of his death, his possessions and his biological characteristics,” to whittle away possibilities from the list of soldiers reported as missing in action from the conflict. But doing so wasn’t quite as easy as it might sound at first blush:

“In an ideal world, there would be a database of the missing and I could conduct a preliminary search based on his height, his dental pattern, his age and his ethnicity. Unfortunately, these data only reside within the individual military records stored in the US National Archives. This meant I needed to determine a short list of possible soldiers and request their records.”

Silverstein knew he could begin with a simple assertion: since Lupo was buried in the same unmarked grave as the unidentified remains, “it was an easy assumption that they died at approximately the same time, July 21 1918, and in about the same location."

From there, Silverstein referred to military maps of the campaign to triangulate which which regiments were in the area where the bodies were found. Having narrowed it down to “to hundreds of MIAs,” Silverstein then had to rely on what the body was buried with to do the rest of the identification:

“The main clues were two buttons on his uniform, one stated “WA” and the other had a “2” and a “D” on it split between two crossed rifles. I discovered that this meant: I-02 had been a member of the Washington State national guard, 2nd regiment, company D, before they were nationalised into the AEF.”

Silverstein used that information, as well as a medal the body had from a 1916 campaign against Mexico, to narrow things down to four men from Company D. Obtaining their records from the National Personnel Records Center, Silverstein used height and dental records to determine that the body recovered belonged to Pfc. Charles McAllister.

As a final step to ensure he was correct, Silverstein contacted Beverly Dillon for mitochondrial DNA analysis. McAllister was Dillon’s great uncle, and she even had the final letter McAllister had sent before being shipped out to France.

“Beverly’s mitochondrial DNA matched Pfc. McAllister," Silverstein writes. “This gave me enough statistical data to show that it was impossible for the remains to belong to anyone else.” More than a century after he was struck down in combat, McAllister can now be laid to rest with military honors in his hometown of Seattle.

Michael Natale is the news editor for Best Products , covering a wide range of topics like gifting, lifestyle, pop culture, and more. He has covered pop culture and commerce professionally for over a decade. His past journalistic writing can be found on sites such as Yahoo! and Comic Book Resources , his podcast appearances can be found wherever you get your podcasts, and his fiction can’t be found anywhere, because it’s not particularly good. 

.css-cuqpxl:before{padding-right:0.3125rem;content:'//';display:inline;} Pop Mech Pro .css-xtujxj:before{padding-left:0.3125rem;content:'//';display:inline;}

Fourth-Dimensional Aliens Could Be Spying On Us

Two Long-Lost Euripedes Works Found in Egypt

Aliens May Possess a Form of Consciousness

Japan Is the World’s Top Hotspot for UFO Sightings

The Army is Building the Next Tiltrotor Aircraft

Are Underwater UFOs an Imminent Threat?

U.S. Not Ready for a Major War: Report

New Clues Emerge About an Ancient Lost City

First F-16s Arrive in Ukraine

A WWII Vet’s Wild Ride on Top of an Airplane Tail

This Laser Could Unlock Interstellar Travel

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Healthcare (Basel)

Use of Advanced Artificial Intelligence in Forensic Medicine, Forensic Anthropology and Clinical Anatomy

Andrej thurzo.

1 Department of Stomatology and Maxillofacial Surgery, Faculty of Medicine, Comenius University in Bratislava, 81250 Bratislava, Slovakia

2 Department of Simulation and Virtual Medical Education, Faculty of Medicine, Comenius University, Sasinkova 4, 81272 Bratislava, Slovakia; [email protected]

3 forensic.sk Institute of Forensic Medical Analyses Ltd., Boženy Němcovej 8, 81104 Bratislava, Slovakia; ks.abinu@1suneb (R.B.); [email protected] (N.M.); moc.liamg@cavokp (P.K.)

Helena Svobodová Kosnáčová

4 Department of Genetics, Cancer Research Institute, Biomedical Research Center, Slovak Academy Sciences, Dúbravská Cesta 9, 84505 Bratislava, Slovakia

Veronika Kurilová

5 Faculty of Electrical Engineering and Information Technology, Slovak University of Technology, Ilkovičova 3, 81219 Bratislava, Slovakia; [email protected]

Silvester Kosmeľ

6 Deep Learning Engineering Department at Cognexa, Faculty of Informatics and Information Technologies, Slovak University of Technology, Ilkovičova 2, 84216 Bratislava, Slovakia; ks.abuts@lemsokx

Radoslav Beňuš

7 Department of Anthropology, Faculty of Natural Sciences, Comenius University in Bratislava, Mlynská dolina Ilkovičova 6, 84215 Bratislava, Slovakia

Norbert Moravanský

8 Institute of Forensic Medicine, Faculty of Medicine, Comenius University in Bratislava, Sasinkova 4, 81108 Bratislava, Slovakia

Peter Kováč

9 Department of Criminal Law and Criminology, Faculty of Law Trnava University, Kollárova 10, 91701 Trnava, Slovakia

Kristína Mikuš Kuracinová

10 Institute of Pathological Anatomy, Faculty of Medicine, Comenius University in Bratislava, Sasinkova 4, 81108 Bratislava, Slovakia; [email protected] (K.M.K.); moc.liamg@retnecevididap (M.P.)

Michal Palkovič

11 Forensic Medicine and Pathological Anatomy Department, Health Care Surveillance Authority (HCSA), Sasinkova 4, 81108 Bratislava, Slovakia

12 Institute of Histology and Embryology, Faculty of Medicine, Comenius University in Bratislava, 81372 Bratislava, Slovakia; [email protected]

Associated Data

We fully adhere to Data Availability Statements in section “MDPI Research Data Policies” at https://www.mdpi.com/ethics (accessed on 1 November 2021).

Three-dimensional convolutional neural networks (3D CNN) of artificial intelligence (AI) are potent in image processing and recognition using deep learning to perform generative and descriptive tasks. Compared to its predecessor, the advantage of CNN is that it automatically detects the important features without any human supervision. 3D CNN is used to extract features in three dimensions where input is a 3D volume or a sequence of 2D pictures, e.g., slices in a cone-beam computer tomography scan (CBCT). The main aim was to bridge interdisciplinary cooperation between forensic medical experts and deep learning engineers, emphasizing activating clinical forensic experts in the field with possibly basic knowledge of advanced artificial intelligence techniques with interest in its implementation in their efforts to advance forensic research further. This paper introduces a novel workflow of 3D CNN analysis of full-head CBCT scans. Authors explore the current and design customized 3D CNN application methods for particular forensic research in five perspectives: (1) sex determination, (2) biological age estimation, (3) 3D cephalometric landmark annotation, (4) growth vectors prediction, (5) facial soft-tissue estimation from the skull and vice versa. In conclusion, 3D CNN application can be a watershed moment in forensic medicine, leading to unprecedented improvement of forensic analysis workflows based on 3D neural networks.

1. Introduction

Conventional forensic analysis is based on forensic expert’s manual extraction of information. Forensic expert provides opinions established on medical and other fields of current knowledge combined with personal work experience. This is not only time-consuming, albeit frequently affected by subjective factors that are tough to overcome [ 1 ].

The main purpose of this paper is to analyze and introduce a very promising line of research applicable to forensic anthropology and various traditional sectors of forensic medicine. The application of artificial intelligence (AI) is a new trend in forensic medicine and a possible watershed moment for the whole forensic field [ 1 , 2 , 3 , 4 , 5 , 6 ].

This chapter paper explains basic terminology, principles and the current horizon of knowledge. The methodology chapter presents the novel clinical workflow based on implementing three-dimensional convolutional neural network (3D CNN) algorithms [ 7 , 8 , 9 ]. The input is full head cone-beam computer tomography scans (CBCT) in the Digital Imaging and Communications in Medicine format (DICOM) [ 9 , 10 , 11 , 12 , 13 , 14 ]. The methodology chapter describes technical data preparation for 3D CNN utilization in the following practical aspects from forensic medicine:

• Biological age determination [ 7 , 8 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 ]
• Sex determination [ 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 ]
• Automatized 3D cephalometric landmark annotation [ 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 ]
• Soft-tissue face prediction from skull and in reverse [ 59 , 60 , 61 , 62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 ]
• Facial growth vectors prediction [ 13 , 59 , 78 , 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 , 87 , 88 , 89 , 90 ]

The result of this paper is a detailed guide for forensic scientists to implement features of 3D CNN to forensic research and analyses of their own (in five themes described above). This resulting practical concept—possible workflow shall be useful for any forensic expert interested in implementing this advanced artificial intelligence feature. This study is based on the worldwide review of 3D CNN use-cases that apply to clinical aspects of forensic medicine

This article’s secondary objective is to inspire forensic experts and approximate them to implement three-dimensional convolutional neural networks (3D CNN) in their forensic research in the fields of age, sex, face and growth determination.

1.1. Basic Terminology and Principles in Era of AI Enhanced Forensic Medicine

Artificial intelligence has brought new vigor to forensic medicine, but at the same time also some challenges. AI and forensic medicine are developing collaboratively and advanced AI implementation until now required extensive interdisciplinary cooperation. In the era of big data [ 3 ], forensic experts shall become familiar with these advanced algorithms and understand used technical terms.

For many forensic experts, the current benefits of advanced AI processes are still unknown. For example, automated AI algorithms for skull damage detection from CT [ 91 ] or soft-tissue prediction of a face from the skull [ 66 , 67 , 89 , 92 ] are still a mystery to many outstanding forensic scientists. Enabling them would catapult forensic research to a new era [ 1 ].

A Convolutional Neural Network (CNN) is a Deep Learning algorithm that can take in an input image, assign importance (learnable weights and biases) to various aspects/objects in the image, and differentiate one from the other.

CNN is an efficient recognition algorithm that is widely used in pattern recognition and image processing. It has many features such as simple structure, less training parameters and adaptability. CNN is a supervised type of Deep learning, most preferable used in image recognition and computer vision ( Figure 1 a,b).

( a ) Recognition of objects. Try, using your imagination, to recognize the objects on the three blurred variants of the same anatomical slice. Convolutional Neural Networks (CNNs) work similar to our visual brain when trying to recognize these objects. ( b ) Our recognition of objects on the picture is significantly improved when more layers—slices are added thus providing further context with the 3rd dimension. In the top row is recognizable intersection of the mandible and vertebra and on the lower row is recognizable slice of the face. 3D CNN recognition is similarly improved with providing context of depth.

Compared to its predecessors, the main advantage of CNN is that it automatically detects the crucial features without any human supervision. For example, given many pictures of cats and dogs, it learns distinctive features for each class. CNN is also computationally efficient.

3D CNN is used to extract features in 3 Dimensions or establish a relationship between 3 dimensions. A 3D CNN is simply the 3D equivalent: it takes as input a 3D volume or a sequence of 2D frames (e.g., CBCT scan).

In terms of Neural Networks and Deep Learning: Convolutions are filters (matrix/vectors) with learnable parameters used to extract low-dimensional features from input data. They have the property to preserve the spatial or positional relationships between input data points.

2D CNNs predict segmentation maps for DICOM slices in a single anatomical plane. 3D CNNs address this issue by using 3D convolutional kernels to make segmentation predictions for a volumetric patch of a scan ( Figure 2 ).

The comparison of 2D CNN ( above ) and 3D CNN ( below ). 3D CNN works with 3rd dimension and can reconstruct shapes from the CBCT 2D slides. The sequence of 2D pictures where the 3rd dimension is time, we speak of a common video sequence that can be a subject of 3D CNN analysis too.

In 3D convolution, a 3D filter can move in all 3-directions (height, width, channel of the image). At each position, the element-wise multiplication and addition provide one number. Since the filter slides through a 3D space, the output numbers are also arranged in a 3D space. The output is then 3D data.

The recognition of similar structures from the CBCT is based on their similar opacity on the X-ray classified by the Hounsfield scale. The process of defining ranges for particular tissues is called “thresholding”, which is prior to final the segmentation ( Figure 3 ). Setting different thresholds for segmentation preprocessing step allows segmentation of different structures such as soft tissues (skin, airway, sinuses), nerves (inferior alveolar nerve, dental pulp), bones (mandible, maxilla or cervical vertebras) and many other ( Figure 4 ).

The example of the process of defining ranges for particular visualized tissues is called “thresholding”.

The examples of the segmentation process on the CBCT data based on defining ranges for particular tissues thus defining 3D structures such as airway, nerve canal, face surface or bone structures.

The segmentation of original CBCT data can result in the definition of various 3D structures involved in 3D CNN training, or these 3D structures can serve as anchors for mapping another 3D scan, such as an intraoral optical scan or extraoral scan that includes texture. All these three data sources can be merged, and the 3D CNN network can work with unprecedented data that include wider face regions from face scan or morphological information on teeth and gums ( Figure 5 ).

The example of 3D data augmentation in a sense of mapping another 3D scans on the segmented structures. Facial 3D scan with texture mapped on the segmented face surface from CBCT and intraoral scan of teeth and gums mapped on tooth surfaces from the CBCT. Finally merged into complex set of 3D models. Training of 3D CNN with such a complex 3D virtualized model has never been performed before and is worth a consideration.

1.2. Overview of Used Artificial Intelligence for Forensic Age and Sex Determination

Traditional forensic analyses of age, gender and facial appearance are based on forensic expert manually acquiring information that provides the identification established on expert`s medical and biological knowledge and mathematical calculations [ 93 , 94 , 95 ]. In forensic outputs, the experiences of the investigator subjectivity and fatigue and emotions play a role [ 93 , 94 , 95 ]. To have forensic expert well trained on thousands of skulls of all possible ethnicities, would take a lifetime. Possible bias sourcing from fatigue, limited training dataset, emotional engagement or human calculation error cannot be absolutely eradicated with human forensic expert. Implementation of artificial intelligence (AI) can limit all these mentioned sources of possible bias. Machine learning works based on models that mimic neurons in the brain and can learn from experiences and solve complex problems. It is not influenced by subjective judgment; it does not become tired and does not use emotions and thus can work more efficiently [ 96 , 97 , 98 ].

AI usage is not without risks of undesired side effects. AI may become biased in the same way as a human forensic expert, depending on the source data used for AI training [ 99 ]. Obermeyer at el. found evidence of bias in a healthcare algorithm responsible for 200 million people, which systemically prevented almost 30% of eligible black patients from receiving additional care by giving lower risk scores to black patients than white patients with equal diagnoses [ 100 ].

Many studies in forensic science have been conducted in recent years, and some recent studies are beginning to focus on neural networks [ 101 ]. These studies were mainly aimed at determining the age and sex of postmortem skeletal remains and living people. Age and gender assessment active, used to identify victims, determine criminal liability or identify persons without legal documentation [ 8 , 102 ]. There is considerable interest in accelerating identification procedures, and experts are involved in machine learning in forensic procedures. They use X-ray images [ 103 , 104 , 105 , 106 , 107 , 108 ], MRI images [ 8 , 109 ], photography [ 90 , 110 , 111 ], CT scans [ 112 , 113 , 114 , 115 , 116 , 117 ] of the head or other bones such as the collarbone, femur, teeth, etc. and use databases to teach artificial intelligence to identify people’s age or gender. Pham et al. [ 113 ] examined age using the femur and mandible for neuronal networks. The femur could play a key role in predicting adulthood, especially the shape of the femoral head and bone densitometry. They used 814 whole-body post-mortem computed tomography (CT) scans to obtain results: 619 men, 195 women aged 20 to 70 years. They omitted subjects with fractures. Each CT output was in digital imaging and communication in medicine (DICOM) format [ 11 , 12 ]. The extracted femur and mandible data were preprocessed to create a 3D voxel inserted into a neural network model. Using this approach, the mean absolute error (MAE) of the mandible age identification was 7.07 years, and the MAE calculated from a femur age determination was 5.74 years. The combination of both approaches reached an excellent result—MAE = 5.15 years. CT scans were also used for learning and age determination in a study by Farhadian et al. [ 115 ]. AI determined the age learned from CT scans of 300 subjects aged 14 to 60 years of the canine teeth. In this study, they compared the methodology of neural networks with a regression model. The MAE for neural networks was 4.12 years, and the MAE for the regression model was 8.17 years, which demonstrated the higher accuracy of neural networks. Mauer et al. [ 102 ] aimed to develop a fully automated and computerized method for age estimation based on the knee’s 3D magnetic resonance imaging (MRI). They had 185 coronal and 404 sagittal MR images of Caucasian men aged 13 to 21 years. The best result obtained was a MAE of 0.67 ± 0.49 years and an accuracy of 90.9%. Here it can be seen that the group with a minor age variance more accurately determines the age of the individuals. A similar study was performed by Stern et al. performed a similar study [ 109 ] where 328 MR images were used for learning neural networks and subsequent age detection. Age was reported with a MAE of 0.37 ± 0.51 years for the age range of individuals ≤ 18 years.

Several research teams have tried neural network learning based on X-ray images [ 103 , 104 , 108 ]. Guo et al. [ 103 ] used 10,257 samples of dental orthopantomograms and, similar to Farhadian et al. [ 115 ], compared logistic regression linear models for each legal age limit (14, 16 and 18 years) with the neural network. The results showed that neural networks work better (linear regression models: 92.5%, 91.3% and 91.8% and neural networks: 95.9%, 95.4% and 92.3% success rate for age limits 14, 16 and 18 years). In Stepanovsky et al. [ 105 ] used 976 orthopantomography (662 men, 314 women) of people aged 2.7 to 20.5 years to learn neural networks. The results were very favorable, and the average absolute error (MAE) was below 0.7 years for both men and women. Vila-Blanco et al. [ 106 ] used landmarks on the mandible to search for patterns by neural networks. The age estimate reached an accuracy of 87.8%, and the MAE was only 1.57 years. De Tobel et al. [ 107 ] used panoramic molar panoramic radiographs to estimate age. The accuracy of the results was, on average, MAE = 0.5. Boedi et al. [ 108 ] later conducted a similar study with similar results. Li et al. [ 104 ] used 1875 X-ray images of the pelvis as a basis for evaluating bone age through deep learning. The age of the people whose X-rays were used to teach the model was 10 to 25 years. The performance of the model was MAE = 0.94 years.

More studies modelled gender determination using AI. Bewes et al. [ 42 ] used neural networks for this purpose with a detection accuracy of 95%. However, they trained them on 900 skull scans from CT scans. Oner et al. [ 114 ] achieved the same goal by using CT images of the sternum transmitted to the orthogonal plane for learning neural networks. They used 422 thin sections of thoracic CT scans (213 females, 209 males) with an age range of 27–60 years. The accuracy of gender prediction was 0.906, and the confidence interval of 94%. The success rate was higher than that achieved by linear models. Etli et al. [ 116 ] compared several methods in the study. They used CT scans with sacral and coccyx metric parameters of 480 patients. They used one-dimensional discriminant analysis, linear discriminant functional analysis, sequential analysis of discriminant function and multilayer perceptron neural networks. The maximum accuracy for each method was 67.1% for one-dimensional discriminant analysis, 82.5% for linear analysis of the discriminant function, 78.8% for sequential analysis of the discriminant function, and 86.3% for multilayer perceptron neural networks.

Gender classification was also discussed by Liew et al. [ 111 ]. The maximum accuracy for each method was 67.1% for one-dimensional discriminant analysis, 82.5% for linear analysis of the discriminant function, 78.8% for sequential analysis of the discriminant function, and 86.3% for multilayer perceptron neural networks. Gender classification was also discussed by Liew et al. [ 111 ]. They used grayscale images of 200 men and 200 women for analysis. The classification performance reached 98.75% and 99.38% in the facial databases SUMS and AT&T. To estimate the sex of infants in the study of Ortega et al. [ 110 ] used 2D photographs of the ilium of 135 individuals aged 5 months to 6 years were used. The accuracy was 59% compared to 61% for the specialist. In addition, Porto et al. [ 88 ] sought to determine the legal age of offenders at 14 and 18 years as Guo et al. [ 103 ]. They based on a database of photographs of 18,000 faces of men and women based on photo anthropometric indices from cephalometric landmarks marked and checked by forensic experts. The accuracy of age determination by neural networks was 0.72 with an age interval of 5 years and for the estimation of the age group higher than 0.93 and 0.83 for the threshold values of 14 and 18 years.

It is almost unbelievable how accurately neural networks can determine age or gender compared to commonly used methods. Therefore, we emphasize their use in forensic practice [ 9 , 46 , 50 , 117 ].

Regarding the Skeletal age estimation for forensic purposes, we consider ourselves useful for the direction of the 3D CNN on particular areas of the head and neck. Various experts published research on age estimation by measuring open apices in teeth, stage of teeth eruption, frequently of third molars or canine tooth/pulp ratio [ 6 , 17 , 18 , 20 , 21 , 23 , 24 , 25 , 27 , 29 , 31 ]. In general, teeth are frequently used for age assessment, but they are not the only structures in the skull to be considered. It is known that the shape of the frontal sinus can be an essential tool in personal forensic identification and is linked together with the cranial base to growth changes that can be evaluated [ 6 , 118 ]. Another typical location for skeletal age assessment in the head and neck X-ray diagnostics region is the stage of cervical vertebrae maturation [ 23 , 119 ]. Deep learning has been already implemented in this area [ 83 ]. Extensive research is published regarding skeletal age expert estimation Pinchi et al. [ 120 , 121 , 122 , 123 , 124 , 125 , 126 ] mainly combines dental and skeletal findings. If the 3D CNN fails to identify these valuable areas, we still have the opportunity to direct the focus on these areas.

Regarding forensic medico-legal aspects, the perspective on natural development estimated by AI algorithms is always relevant, especially in the situation of trauma or other damage that conflicted with this estimated development. AI is now used to evaluate CT scans of lungs and to predict the deterioration of COVID-19 patients in the emergency department [ 127 , 128 , 129 ].

In this case, 3D CNN algorithms can automatically evaluate not only hard-tissue structures and search for inapparent damage that could have been responsible for a sudden death incident [ 91 , 130 ].

1.3. Artificial Intelligence Implementation in 3D Cephalometric Landmark Identification

Analysis of complex cranial and facial structures is a domain of orthodontics. Historically they are fundamental for proper treatment planning, and they represent lines, angles, planes on the axilla-facial structures identifiable, especially on the X-ray (typically lateral X-ray). There is massive research regarding cephalometric parameters and their values. Observer defines the points, and their interobserver error are the main weakness of cephalometric analysis ( Figure 6 ). Anthropometry in Forensic Medicine and Forensic Science is frequently used for sex and biological age determination and other purposes [ 129 , 130 , 131 ].

Example of 3D cephalometric analysis where orthodontist identifies more than 50 points and the hard- and soft-tissues analyzed. Humans chose these points as the most reproducible on X-ray. These might not be ideal representatives of head and neck structures linked with biological ageing or sexual dimorphism.

As the various cephalometric parameters (angles, ratios and distances) were well researched, and some are proven to be related to age, sex or growth, they are a frequent springboard for many research studies focused on facial parameters. Implementation of AI in cephalometric analysis has been published [ 132 , 133 , 134 , 135 , 136 ]. The question is whether the 3D CNN trained networks will find even better regions and soft- and hard-tissue features on CBCTs when autonomously searching for links between voxel structures and the age or sex. Either way, the reliable automatized 3D cephalometric algorithm precisely identifying particular points with extreme repeatability would be a helpful tool not intended to replace humans in cephalometric points identifications. However, the human error is impossible to cancel completely as the interobserver error.

1.4. Artificial Intelligence Implementation in Soft-Tissue Face Prediction from Skull and Vice Versa

Reconstruction of the face from the skull is an age-old desire of forensic experts. Current methods of not implementing AI are very limited. Prediction of soft tissues according to the hard tissues of the skull and vice versa can be significantly improved upon big-data training of 3D CNN with supplementary metadata about age, sex, BMI or ethnicity. New algorithms to perform facial reconstruction from a given skull has forensic application in helping the identification of skeletal remains when additional information is unavailable [ 29 , 64 , 66 , 67 , 68 , 69 , 70 , 72 , 73 , 85 , 86 , 88 , 89 , 92 , 137 ]. Implementation of 3D CNN can also unintentionally open pandora box of guided improving the morphology of the facial soft-tissues. From a socio-psychological standpoint, this is regarded as an important therapeutic goal in modern orthodontic treatments. Currently, many of the algorithms implemented in commercially available software present ability to predict profile changes grounded on the incorrect assumption that the amount of movement of hard-tissue and soft-tissue has a proportional relationship [ 82 ].

The beauty industry has seen rapid growth in multiple countries, and due to its applications in entertainment, the analysis and assessment of facial attractiveness have received attention from scientists, physicians, and artists because of digital media, plastic surgery, and cosmetics. An analysis of techniques is used to assess facial beauty that considers facial ratios and facial qualities as elements to predict facial beauty [ 81 , 82 , 138 , 139 , 140 ]. A popular and famous free app using AI is FaceApp, which uses neural networks to enhance, age or otherwise change 2D digital photos of users uploading them using this application ( Figure 7 ). Using the 3D CNN approach was not yet implemented despite iPhones having a 3D lidar scanner to acquire a 3D soft-tissue scan of the user’s face. From a forensic aspect, this era of digital 2D face manipulation brought deep-fake videos and images. Detecting manipulated facial images and videos is an increasingly important topic in digital media forensics [ 118 , 141 ]. Any face can be used in the fake video, or unlimited numbers of nearly authentic pictures, including fake social media profiles, can be created. AI is used in forensic evaluation for facial forgery detection and manipulated region localization [ 118 ].

Example of CNN use of the FaceApp AI application to render the face mapped on CBCT to look younger or older. The algorithms changed just the texture and not the 3D mask, however this is probably only a matter of time. 2D face morphing based on AI or face swapping in popular videos are available and popular already a couple of years. Original face is the 2nd one.

This paper most complex AI application is the final 5th theme—“Facial growth vectors prediction”. The authors of this paper addressed it for various reasons. Firstly, it is fundamentally different from the first four themes. Secondly, it requires the most complex implementation of AI strategies. To our knowledge, this is only the second paper in the world that handles the problem of facial growth prediction with ML methods and absolutely the first paper to consider a 3D CNN for facial growth predictions.

Prediction of natural growth is compared to typically forensic topics such as human remains reconstruction and identification or age and sex determination rather less familiar topic. Mainly because despite numerous research attempts to predict facial growth, a satisfactory method has not been established yet, and the problem still poses a challenge for medical experts [ 142 , 143 , 144 ]. Predicting natural growth and later ageing is relevant for orthodontic therapy planning and from a forensic aspect. Any damage to the head and neck region that would affect otherwise natural growth or simple ageing could be evaluated. The effect of such a trauma could be in the future forensically quite accurately evaluated.

In 1971 Hirschfeld and Moyers published an article named “Prediction of craniofacial growth: the state of the art” [ 144 ]. This was one of the first attempts for facial growth predictions. The authors concluded that there are many reasons why they fail to predict craniofacial growth, and some they named persisted until today. They expressed doubts that we have not always measured the right thing. They also pointed out the lack of biological meaning for many traditional cephalometric measures. They have also pointed to the heritability of attained growth in the face and predicted the future importance of craniofacial genetics. The future that comes proved them correct in many aspects. Since these first attempts to predict the facial growth direction over half of a century ago, we did not become much better in facial growth prediction [ 142 ]. The complexity of the problem is challenging.

The only study that was focused on the prediction of the facial growth direction with Machine Learning methods and has been published so far is a paper with its pre-print [ 90 , 145 ] from 2021 by Stanislaw Kazmierczak et al. The outcomes of this paper are not impressive regarding facial growth prediction, albeit inspiring in the method of evaluation. The authors of this novel paper [ 94 ] performed feature selection and pointed out the attribute that plays a central role in facial growth. Then they performed data augmentation (DA) methods. This study is discussed in more detail later in this paper.

2. 3D Convolutional Neural Networks and Methods of Their Use in Forensic Medicine

2.1. hardware and software used.

CBCT scans analyzed for this paper were made on one machine: i-CAT™ FLX V17 with the Field of View (FOV) of 23 cm × 17 cm with technical parameters and settings Table 1 .

Full-head CBCT scans were mate with i-CAT™ FLX V17 with these settings.

Medical software used for DICOM data processing and analysis was Invivo™ 6 from Anatomage Inc., Silicon Valley, Thomas Road Suite 150, Santa Clara, CA 95054, USA.

Software for the AI solution base we have used the Python programming language along with 3 deep learning libraries—TensorFlow 2, PyTorch and MONAI. As for the hardware, the whole AI system is powered by multiple GPUs.

2.2. Main Tasks Definitions

Task 1—Age estimation from whole 3D CT scan image

Definition: the task is to estimate the approximate age of a person from a whole head 3D CBCT scan

Proposed method: build regression model represented by a 3D deep neural network that has the current state of the art network architecture as a backbone

Metrics: Mean Absolute Error (MAE) and Mean Squared Error (MSE) (see Section Evaluation)

Task 2—Sex classification from thresholded soft and hard tissues

Definition: the task is to classify input 3D CBCT scans (whole head or experimentally segmented parts) into one of 2 predefined categories—female and male

Proposed method: build classification model represented by 3D deep neural network based on convolutional layers and outputs class probabilities for both targets

Metrics: Accuracy and Confusion Matrix (CM) (other metrics such as precision, recall and F1 score will be evaluated in a later phase)

Task 3—Automatization of cephalometric measurements

Definition: the task is to create an automated system able to tag cephalometric landmarks on whole head 3D CT scan

Proposed method: build object detection model based on 3D neural network that estimates cephalometric measurements automatically

Task 4—Soft-tissue face prediction from skull and vice versa

Definition: the task is to create an automated system that predicts the distance of the face surface from the bone surface according to the estimated age and sex. 3D CNN to be trained on whole-head CBCTs of soft-tissue and hard-tissue pairs. *CBCTs with trauma and other unnatural deformations shall be excluded.

Proposed method: build a generative model based on Generative Adversarial Network that synthesizes both soft and hard tissues

Metrics: the slice-wise Frechet Inception Distance (see Section Evaluation)

Task 5—Facial growth prediction

Definition: the task is to create an automated system that predicts future morphological change in defined time for the face’s hard- and soft tissues. This shall be based on two CBCT input scans of the same individual in two different time points. The second CBCTs must not be deformed with therapy affecting morphology or unnatural event. This already defines the extremely challenging condition. There is a high possibility of insufficient datasets and the necessity of multicentric cooperation for successful training of 3D CNN on this task.

Proposed method: In this final complex task, the proposed method builds on previous tasks. We strongly recommend adding metadata layers on gender, biological age and especially genetics or letting the CNN determine them by itself. We suggest disregarding the established cephalometric points, lines, angles and plains as these were defined in regards to lateral X-ray, emphasising good contrast of the bone structures with high reproducibility of the point and not necessarily with focus on particular structures most affected by growth. We suggest letting3D CNN establish its observations and focus areas.

We also suggest allowing 3D CNN analysis of genetic predisposition in a smart way: by analysis of possibly CBCT of the biological parents or preferably non-invasive face-scan providing at least facial shell data.

2.3. The Data Management

The processing of data in deep learning is crucial for the sufficient result of any neural network. Currently, most of the implementations depend on the dominant model-centric approach to AI, which means that developers spend most of their time improving neural networks.

For medical images, various preprocessing steps are recommended. In most cases, the initial steps are following ( Figure 8 ):

• Loading DICOM files—the proper way of loading the DICOM file ensures that we will not lose the exact quality
• Pixel values to Hounsfield Units alignment—the Hounsfield Unit (HU) measures radiodensity for each body tissue. The Hounsfield scale that determines the values for various tissues usually ranges from −1000 HU to +3000 HU, and therefore, this step ensures that the pixel values for each CT scan do not exceed these thresholds.
• Resampling to isomorphic resolution—the distance between consecutive slices in each CT scan defines the slice thickness. This would mean a nontrivial challenge for the neural network. The thickness depends on the CT device setup, and therefore it is necessary to create equally spaced slices.
• [Optional] Specific part segmentation—each tissue corresponds to a specific range in the Hounsfield units scale, and in some cases, we can segment out specific parts of the CT scan by thresholding the image.
• Normalization and zero centering—these two steps ensure that the input data that are feed into the neural network are normalized into [0, 1] interval (normalization) and are zero centered (achieved by subtracting the mean value of the image pixel values).

For medical images, there are various preprocessing steps that are recommended.

Preprocessing the image dataset before feeding the CNN or other classifiers is important for all imaging modalities. Several preprocessing steps are recommended for the medical images before they are fed as input to the deep neural network model, such as (1) artefact removal, (2) normalization, (3) slice timing correction (STC), (4) image registration and (5) bias field correction. While all the steps (1) to (5), help in acquiring reliable results, STC and image registration are very important in the case of 3D medical images (especially nMR and CT images). Artefact removal and normalization are the most performed preprocessing steps across the modalities [ 146 ].

2.4. Dataset Specification

This study comprises approximately 500 iCAT CBCT scans of human heads. Each CBCT scan has the spatial resolution of 768 × 768 pixels and the default device pixel spacing is [0.3 × 0.3 × 0.3] millimeters.

The subjects are split by sex, with the ratio of 6:4 for female/male ranging from 8 to 72 years. The majority (90%) of the subjects are between 18 and 36 years.

These dataset parameters were used in suggested considerations for 3D CNN applications concepts stated in Section 2.2 Main tasks definitions.

2.5. Deep Learning Approach

2.5.1. age estimation using 3d deep neural networks.

In recent research [ 7 , 14 ] AI-based age estimation has proven to be a successful competitor to classical approaches from forensic medicine. The aim of this study is to create an automated system for age estimation from 3D cranial CT scans. There is an expectation that particular parts of the skull have a decisive impact on the final prediction, and therefore we propose a solution that includes two stages:

Age estimation from dense tissue layer—we use whole skull CT scan as an input into the 3D convolutional neural network, which would serve as a regression model that estimates the continuous values of age for each CT scan separately.

[Experimental] Visualization of network activations that represent regions of interest—neural network’s intermediate layers often serve as a an excellent explaining tool in order to find visual explanation heat maps [ 113 ] that highlight regions that affect neural network the most.

As for the specific neural network architecture, we derive the backbone part from the current state of the art research. We primarily consider the EfficientNet [ 147 ] and DenseNet [ 148 ] with their implementations adapted to 3D inputs.

Both architectures base includes convolutional layers that serve as feature extraction blocks to obtain specific indicators from input x represented as a loaded DICOM image. These extracted feature maps are then forwarded to a fully-connected layer that outputs the single age estimation value:

where CL is an intermediate block consisting of convolutional layers, FC is a fully-connected top part of the network that outputs a single floating-point value.

2.5.2. Sex Classification Using 3D Deep Neural Networks

The determination of sex from human remains is a challenging task in various fields such as archeology, physical anthropology and forensics because there is no proven method that exactly leads to correct classification.

The use of AI in this field is highly desirable as manual determination is often very complex and time-consuming. The objectiveness of the deep learning approach can also eliminate human bias leading to reliable software products.

The sex classification is carried out similarly to the previous age estimation approach, but this task´s objective is to classify the final outputs from the neural network into 2 classes—female and male. For this purpose, we use the softmax activation function as a last operation to obtain class probabilities for both targets. The computation is following:

where CL and FC represent the convolutional and fully-connected blocks of the neural network.

The experimental part would include the input x consisting of 2 separate inputs—one will be the segmented skull and the other will be the segmented soft tissue (skin) which is achieved by setting different thresholds for segmentation preprocessing step.

2.5.3. Automatization of Cephalometric Analysis

The cephalometric analysis aims to set landmarks of CT(CBCT) scans which serve as an important factor in the alignment of a skull. These measurements can also be used as surgery planning parameters or pre-and post-surgery comparisons [ 149 , 150 ].

The idea behind this approach is to use 3D convolutional neural networks for fully automated cephalometric analysis. Networks aim to output probabilistic estimations for each cephalometric landmark and then create a projection of these estimations into a real skull CT scan ( Figure 9 ).

Pipeline from pre-processed CBCT scans to prediction on 3D CNN.

Two approaches come into consideration:

• Landmarks estimation in whole CT scan image—in this approach, the probability estimation for all landmarks is assigned for each pixel in the CT scan
• Landmarks estimation for selected regions of interest—assuming that each landmark corresponds to a specific area we could add another preprocessing step—slice cut where each slice would be a template-based region fed into a neural network. We can determine the expected landmark detection for each slice independently, which should help in the final model performance

2.5.4. Neural Networks Architectures and Clinical Data Pre-Processing

Recently, CNNs have been successfully applied in widespread medical image analysis and achieved significant benefits [ 9 , 59 , 115 , 141 , 151 ]. We investigated the design of a 3D CNN with backbones based on Resnet, MobileNet, and SqueezeNet models, which have proven to be the most efficient and widely used in various applications. One of the preferable architectures was based on 3D Resnet34 for the mandible segmentation in research of Pham et al. 2021 [ 113 ].

We have considered various approaches:

• Use whole 3D CT scan as an input into the neural network and output one value for age estimation as floating value and one for sex classification as a binary value.
• Segment out the mandible and use it as input into the neural network. Output is the same as in the previous task.
• (experimental) Use a whole 3D CT scan to input into the neural network and output multiple values representing specific skull features (as discussed at the meeting last week). Then use these values as an input into another machine learning model to estimate age and gender.

Suppose we take an example of mandible segmentation from DICOM. The first step is to have DICOM files loaded and then, added any missing metadata; particularly, the slice thickness, that is, the pixel size in the Z direction, which was obtained from the DICOM file. The unit of measurement in CBCT scans is the Hounsfield Unit (HU), which is a measure of radiodensity. Thus, HU shall be converted to pixel values. Subsequently, it shall be resampled to an isomorphic resolution to remove the scanner resolution. The slice thickness refers to the distance between consecutive slices (when viewing a 3D image as a collection of 2D slices) and varies between scans.

The final preprocessing step is bone segmentation and pixel normalization. Mandible bone extraction is complex because the surrounding bone has to be removed. An image binary thresholding and morphological opening operation for each slice shall be applied.

The morphological opening operation is an essential technique in image processing, achieved by erosion and the dilation of an image. This technique helps to remove small objects while retaining more significant parts from an image. To obtain the mandible bone part, the largest areas after morphological opening shall be kept. Finally, all the slices shall be stacked together to obtain the mandible voxels.

2.6. Evaluation

All approaches are evaluated in a classical machine learning manner—the dataset is split into three parts train, validation and test split. The test split mainly serves as a benchmarking set in order to compare our results with other approaches.

2.6.1. Regression Models

When dealing with regression models in the deep learning field, we usually take into consideration two main regression metrics—Mean Absolute Error (MAE) and Mean Squared Error (MSE). Both metrics calculate the error between predicted y and ground truth labels denoted as y.

MAE is defined as the mean of the sum of absolute differences between y and ŷ for each pixel:

while MSE is defined as mean of the squares of the errors, where error is defined as difference between y and ŷ:

the regression tasks are primarily related to Task 1—age estimation and Task 3—automated cephalometric analysis.

2.6.2. Classification Models

In order to evaluate the classification task, which in our case is represented by Task 2—sex classification, we need to consider the current distribution of male and female samples in our dataset. As the distribution is approximately 6:4 (almost equal), we can calculate the overall accuracy and corresponding confusion matrix (CM). In the later phase, we can also test other metrics such as precision, recall or F1 score.

The calculation of accuracy is defined just as the number of correct predictions divided by the total number of predictions. More interesting for use would be the CM. It is a tabular visualization of a model prediction for each class separately ( Figure 10 ).

Confusion matrix for 2 classes image classification—Cat and Non-Cat. Each row corresponds to the predicted class from neural network output. In case of a class Cat 90 samples were correctly classified as Cat and 60 samples were incorrectly classified as Non-Cat.

3. Resulting Summary of Proposed Approach for Utilization of 3D CNN in Investigated Aspects of Forensic Medicine

This chapter is presenting summary outcome from the detailed research in previous sections of this paper. Investigation of 3D CNN modalities, their features, advantages and disadvantages and also clinical requirements for implementation in the field of forensic medicine has led to these proposed designs (guide) of future forensic research based on 3D CNN analyses.

Table 2 presents condensed summary of recommended approach for 3D CNN implementations in various forensic topics. Expected input data is the minimal dataset of 500 full-head CBCT scans, described in more detail in previous sections.

Guide of recommended designs for 3D CNN implementations in various forensic topics.

1 Method and metrics are not proposed from the current state of knowledge for Facial growth prediction and need further consideration upon clinical experience from 3D CNN applications.

4. Discussion

The authors of this paper have no doubts that 3D CNN, as another evolutionary step in advanced AI, will be with practical implementation a watershed moment in forensic medicine fields dealing with morphological aspects.

With considered data input as CT or CBCT (DICOM), the implementation of 3D CNN algorithms opens unique opportunities in areas of:

• Biological age determination
• Sex determination
• ○ 3D cephalometric analysis of soft and hard tissues
• ○ 3D face prediction from the skull (soft-tissues) and vice versa
• ○ Search for hidden damage in post-mortem high-resolution CT images
• ○ Asymmetry and disproportionality evaluation
• ○ Hard-tissue and soft tissue growth
• ○ Aging in general
• ○ Ideal face proportions respecting golden ratio proportions
• ○ Missing parts of the skull or face
• ○ 3D dental fingerprints for identification with 2D dental records

First clinical applications of 3D CNN have shown [ 91 , 113 , 115 , 126 , 150 ] that the algorithms can be successfully used in CT analysis and identifications of specific diseases such as Alzheimer or COVID19 as these have a specific representation on the X-ray. With a high probability bordering on certainty, the future development of advanced 3D CNN will result in sophisticated automatized algorithms processing 3D diagnostic data similarly to the trained human eye of the forensic expert. These algorithms will automatically process 3D diagnostic data such as CT or NMR, searching for patterns they were trained to see. They will recognize unseen details of hidden damage or representations of rare diseases when trained to do so. In the next level, they will approximate the finding to become an ultimate autopsy tool for even unknown diseases [ 36 , 113 , 126 , 152 ].

The limitation of this paper is that practical examination of the proposed directions for 3D CNN implementations will require some time. Currently, there are many different 3D CNN in development, and actually, this is where most of the research activity is carried out [ 151 , 153 , 154 , 155 ].

Another limitation of this study is the high level of dynamics of research and development in this field of advanced AI implementations. The velocity in training the 3D CNN is high, and it is possible that a better approach can be recognized in the process.

Interesting limitation of 3D CNN usage is the known fact [ 99 ] the any AI may become biased in the same way as a human forensic expert does and not only in the context of the criminal trial. This depends on the source data used for AI training [ 99 ] and is elaborated in more context in Section 1.2 . On the other hand, in many forensic cases we need to achieve highest probabilities on the boundary with certainty. Here a respected and internationally recognized algorithm might become a useful tool for achieving an unprecedented levels of probability superior to human evaluation. However, this development is a possibility, not certainty.

The final limitation of implementing the suggested designs for 3D CNN implementation for forensic researchers is the physical and legal availability of big data necessary for 3D CNN training. This can be solved with multicentric cooperation.

There already exist many CNN processing DICOM data and are available for use [ 11 , 12 , 14 ]. Researchers this year have already achieved significant milestones in multiclass CBCT image segmentation for orthodontics with Deep Learning. They trained and validated a mixed-scale dense convolutional neural network for multiclass segmentation of the jaw, the teeth, and the background in CBCT scans [ 153 ]. This study showed that multiclass segmentation of jaw and teeth was accurate, and its performance was comparable to binary segmentation. This is important because this strongly reduces the time required to segment multiple anatomic structures in CBCT scans.

In our efforts, we have faced the issue of CBCT scan distortion caused by metal artefacts (mostly by amalgam dental fillings). Fortunately, a novel coarse-to-fine segmentation framework was recently published based on 3D CNN and recurrent SegUnet for mandible segmentation in CBCT scans. Moreover, the experiments indicate that the proposed algorithm can provide more accurate and robust segmentation results for different imaging techniques compared to the state-of-the-art models with respect to these three datasets [ 156 ].

As there already exists a fully automated method for 3D individual tooth identification and segmentation from dental CBCT [ 154 ], these algorithms can be combined.

The most complex area covered by this paper is a 3D prediction of growth and in a wider perspective of ageing. It is known that this process is laden with various variables including hormonal (sex) [ 142 , 143 , 157 , 158 , 159 ] and functional aspects (bad habits) [ 160 , 161 , 162 ], as well as genetics [ 163 , 164 , 165 , 166 ].

The only published study focused on predicting the facial growth direction with the implementation of Machine Learning methods is from 2021 Kazmierczak et al. 2021 [ 90 , 145 ]. The outcomes of this paper are limited in regards to facial growth prediction. The authors of this original paper did feature selection and pointed the attribute that plays a central role in facial growth. Then they performed data augmentation (DA) methods.

The principal weakness of this study is not the method but probably the input. The authors used only 2D lateral X-rays of various qualities and sizes. In addition, the evaluation was performed only in one 2D projection. The researchers focused on the angle between the Sella—Nasion line and the Mandibular plane formed by connecting the point gonion to gnathion at the inferior border of the mandible. They engaged an orthodontic expert to identify approximately 20 characteristic anatomic landmarks on LC to assess a subject. These were assessed manually on the lateral cephalogram. Some of the landmarks define angles which, from the clinical perspective, have special significance. As far as facial growth direction is concerned, there are no standardized measurements available in the literature to evaluate. The focus of supervised ML with a concentration on established cephalometric parameters might be wrong. It is the fact that they were originally chosen as well distinguished points on lateral X-ray with a priority of high reproducibility. So as considered by Hirschfeld and Moyers more than 50 years ago, we might be looking in the wrong places. Prediction of the change of SN/MP also oversimplifies the problem. The questions from the past remain, and facial growth prediction remains a complex mystery. The application of 3D CNN on this most complex task is described in more technical details and discussed later in the context of addressing other variables such as biological age, sex and genetics.

5. Conclusions

In conclusion, we can assume that the 3D CNN, as an advanced AI feature, will shift the paradigm in all areas researched in this paper. Forensic experts are now guided to step into the era of artificial intelligence as a helpful tool for research and possibly even future routine forensic analyses. Proposed methods and metrics for 3D CNN application on particular forensic topics (Biological age determination, Sex determination, 3D cephalometric analysis and Face prediction from skull), summarized in resulting Table 2 , can be used as the initial guide. Forensic 3D reconstructions using artificial intelligence will be new, exciting and practically usable methods.

The implementation of advanced AI still requires interdisciplinary cooperation, albeit, with understanding, it can be used to crack unsolved mysteries. It definitely is not a trend that can be ignored.

Acknowledgments

We acknowledge technological support of Cognexa software company for support and digital dental lab infrastructure of 3Dent Medical s.r.o company as well as dental clinic Sangre Azul s.r.o.

Abbreviations

Author Contributions

Conceptualization, A.T., H.S.K., V.K., S.K., R.B., N.M., P.K., K.M.K., M.P. and I.V.; methodology, A.T., H.S.K., V.K., S.K., R.B., N.M., P.K., K.M.K., M.P. and I.V.; software, A.T., H.S.K., S.K. and I.V.; validation, A.T., H.S.K., V.K., S.K. and I.V.; formal analysis, A.T., H.S.K., V.K., S.K. and I.V.; investigation, A.T., H.S.K., V.K., S.K. and I.V.; resources, A.T.; data curation, A.T.; writing—original draft preparation, A.T., H.S.K., V.K., S.K., R.B., N.M., P.K., K.M.K., M.P. and I.V.; writing—review and editing, A.T., H.S.K., V.K., S.K., R.B., N.M., P.K., K.M.K., M.P. and I.V.; visualization, A.T.; supervision, A.T., H.S.K., S.K. and I.V.; project administration, A.T.; funding acquisition, A.T., H.S.K., S.K. and I.V.; All authors have read and agreed to the published version of the manuscript.

This research was funded by the KEGA grant agency of the Ministry of Education, Science, Research, and Sport of the Slovak Republic (Grant No. 081UK-4/2021).

Institutional Review Board Statement

The study was conducted according to the guidelines of the Declaration of Helsinki, and no approval was necessary by the Ethics Committee. Ethical review and approval were waived for this study, due to the fact that no experimental materials or approaches were used.

Informed Consent Statement

Written informed consent was obtained from all subjects involved in the study.

Data Availability Statement

Conflicts of interest.

The authors declare no conflict of interest.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Grab your spot at the free arXiv Accessibility Forum

Help | Advanced Search

Computer Science > Artificial Intelligence

Title: the ai scientist: towards fully automated open-ended scientific discovery.

Abstract: One of the grand challenges of artificial general intelligence is developing agents capable of conducting scientific research and discovering new knowledge. While frontier models have already been used as aides to human scientists, e.g. for brainstorming ideas, writing code, or prediction tasks, they still conduct only a small part of the scientific process. This paper presents the first comprehensive framework for fully automatic scientific discovery, enabling frontier large language models to perform research independently and communicate their findings. We introduce The AI Scientist, which generates novel research ideas, writes code, executes experiments, visualizes results, describes its findings by writing a full scientific paper, and then runs a simulated review process for evaluation. In principle, this process can be repeated to iteratively develop ideas in an open-ended fashion, acting like the human scientific community. We demonstrate its versatility by applying it to three distinct subfields of machine learning: diffusion modeling, transformer-based language modeling, and learning dynamics. Each idea is implemented and developed into a full paper at a cost of less than $15 per paper. To evaluate the generated papers, we design and validate an automated reviewer, which we show achieves near-human performance in evaluating paper scores. The AI Scientist can produce papers that exceed the acceptance threshold at a top machine learning conference as judged by our automated reviewer. This approach signifies the beginning of a new era in scientific discovery in machine learning: bringing the transformative benefits of AI agents to the entire research process of AI itself, and taking us closer to a world where endless affordable creativity and innovation can be unleashed on the world's most challenging problems. Our code is open-sourced at this https URL

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .