Automated Radiological Identification Using 3D Segmentation of the Skull Bone from CT Data

IF 0.9 Q4 RADIOLOGY, NUCLEAR MEDICINE & MEDICAL IMAGING

Forensic Imaging Pub Date : 2026-05-01 Epub Date: 2026-05-14 DOI:10.1016/j.fri.2026.200758

Joel Bottoni , Holger Wittig , Eva Scheurer , Claudia Lenz , Dominique Neuhaus

{"title":"Automated Radiological Identification Using 3D Segmentation of the Skull Bone from CT Data","authors":"Joel Bottoni , Holger Wittig , Eva Scheurer , Claudia Lenz , Dominique Neuhaus","doi":"10.1016/j.fri.2026.200758","DOIUrl":null,"url":null,"abstract":"<div><h3>Background</h3><div>A reliable method for the identification of deceased individuals is radiological identification through comparison of antemortem (AM) and postmortem (PM) images of anatomical structures in the head region, such as teeth or paranasal sinuses. However, this method is examiner-dependent and time-consuming. The aim of this study was to automate the comparison of image data of the entire skull.</div></div><div>A reliable method for the identification of deceased individuals is radiological identification through comparison of antemortem (AM) and postmortem (PM) images of anatomical structures in the head region, such as teeth or paranasal sinuses. However, this method is examiner-dependent and time-consuming. The aim of this study was to automate the comparison of image data of the entire skull.</div><div><h3>Methods</h3><div>The skull bone was automatically segmented in 30 matching AM and PM CT datasets and 18 additional AM CT datasets using the open-source tool TotalSegmentator. Using a custom Python script, the segmented AM and PM data were registered with the image registration tool FSL FLIRT, and the degree of overlap was calculated based on the precision score (0 = no overlap, 1 = maximum overlap). The script compared each PM case with all 30 AM cases as well as the additional 18 AM cases, computed the individual overlap values, and classified the case with the highest overlap as a match.</div></div><div>The skull bone was automatically segmented in 30 matching AM and PM CT datasets and 18 additional AM CT datasets using the open-source tool TotalSegmentator. Using a custom Python script, the segmented AM and PM data were registered with the image registration tool FSL FLIRT, and the degree of overlap was calculated based on the precision score (0 = no overlap, 1 = maximum overlap). The script compared each PM case with all 30 AM cases as well as the additional 18 AM cases, computed the individual overlap values, and classified the case with the highest overlap as a match.</div><div><h3>Results</h3><div>In 28 out of 30 PM cases, the correct AM case was identified as the match. The Dice scores of the correct matches ranged from 0.78 to 0.99, while the scores of the incorrect matches ranged from 0.37 to 0.93.</div></div><div>In 28 out of 30 PM cases, the correct AM case was identified as the match. The Dice scores of the correct matches ranged from 0.78 to 0.99, while the scores of the incorrect matches ranged from 0.37 to 0.93.</div><div><h3>Conclusions</h3><div>The results suggest that automated 3D segmentation, registration, and similarity analysis of CT data of the entire skull can reliably, rapidly, and examiner-independently determine the correct match for a PM case. Further cases and optimizations are required to validate the method and to define a threshold value for the precision score.</div></div><div>The results suggest that automated 3D segmentation, registration, and similarity analysis of CT data of the entire skull can reliably, rapidly, and examiner-independently determine the correct match for a PM case. Further cases and optimizations are required to validate the method and to define a threshold value for the precision score.</div>","PeriodicalId":40763,"journal":{"name":"Forensic Imaging","volume":"45 ","pages":"Article 200758"},"PeriodicalIF":0.9000,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Forensic Imaging","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2666225626000898","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2026/5/14 0:00:00","PubModel":"Epub","JCR":"Q4","JCRName":"RADIOLOGY, NUCLEAR MEDICINE & MEDICAL IMAGING","Score":null,"Total":0}

引用次数: 0

Abstract

Background

A reliable method for the identification of deceased individuals is radiological identification through comparison of antemortem (AM) and postmortem (PM) images of anatomical structures in the head region, such as teeth or paranasal sinuses. However, this method is examiner-dependent and time-consuming. The aim of this study was to automate the comparison of image data of the entire skull.

Methods

The skull bone was automatically segmented in 30 matching AM and PM CT datasets and 18 additional AM CT datasets using the open-source tool TotalSegmentator. Using a custom Python script, the segmented AM and PM data were registered with the image registration tool FSL FLIRT, and the degree of overlap was calculated based on the precision score (0 = no overlap, 1 = maximum overlap). The script compared each PM case with all 30 AM cases as well as the additional 18 AM cases, computed the individual overlap values, and classified the case with the highest overlap as a match.

Results

In 28 out of 30 PM cases, the correct AM case was identified as the match. The Dice scores of the correct matches ranged from 0.78 to 0.99, while the scores of the incorrect matches ranged from 0.37 to 0.93.

Conclusions

The results suggest that automated 3D segmentation, registration, and similarity analysis of CT data of the entire skull can reliably, rapidly, and examiner-independently determine the correct match for a PM case. Further cases and optimizations are required to validate the method and to define a threshold value for the precision score.

查看原文本刊更多论文

利用CT数据对颅骨进行三维分割的自动放射学识别

背景鉴定死者个体的可靠方法是通过比较头部区域解剖结构（如牙齿或鼻窦）的死前（AM）和死后（PM）图像进行放射学鉴定。然而，这种方法依赖于考官，而且耗时。本研究的目的是实现整个颅骨图像数据的自动比较。鉴别死者的一种可靠方法是通过比较头部区域解剖结构（如牙齿或鼻窦）的死前（AM）和死后（PM）图像进行放射学鉴定。然而，这种方法依赖于考官，而且耗时。本研究的目的是实现整个颅骨图像数据的自动比较。方法采用开源软件TotalSegmentator对30组匹配的AM和PM CT数据集以及另外18组AM CT数据集进行颅骨自动分割。使用自定义Python脚本，将分割后的AM和PM数据与图像配准工具FSL FLIRT进行配准，并根据精度评分（0 = 无重叠，1 = 最大重叠）计算重叠程度。该脚本将每个PM病例与所有30个AM病例以及额外的18个AM病例进行比较，计算单个重叠值，并将重叠最高的病例分类为匹配。使用开源工具TotalSegmentator对30个匹配的AM和PM CT数据集以及18个额外的AM CT数据集进行颅骨自动分割。使用自定义Python脚本，将分割后的AM和PM数据与图像配准工具FSL FLIRT进行配准，并根据精度评分（0 = 无重叠，1 = 最大重叠）计算重叠程度。该脚本将每个PM病例与所有30个AM病例以及额外的18个AM病例进行比较，计算单个重叠值，并将重叠最高的病例分类为匹配。结果在30例PM病例中，有28例正确的AM病例被识别为匹配。正确匹配的骰子得分在0.78到0.99之间，而错误匹配的得分在0.37到0.93之间。在30个PM病例中，有28个正确的AM病例被确定为匹配。正确匹配的骰子得分在0.78到0.99之间，而错误匹配的得分在0.37到0.93之间。结论全颅骨CT数据的自动三维分割、配准和相似度分析可以可靠、快速、独立于检查者确定PM病例的正确匹配。需要进一步的用例和优化来验证该方法并定义精度分数的阈值。结果表明，全颅骨CT数据的自动三维分割、配准和相似性分析可以可靠、快速、独立于检查者确定PM病例的正确匹配。需要进一步的用例和优化来验证该方法并定义精度分数的阈值。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Forensic Imaging RADIOLOGY, NUCLEAR MEDICINE & MEDICAL IMAGING-

CiteScore

2.20

自引率

27.30%

发文量