Artificial Intelligence (AI), Hanging, Postmortem CT (PMCT)

Abstract

Background

Estimation of the postmortem interval (PMI) is important in forensic practice. In recent years, the use of postmortem CT (PMCT) has been increasing in the field of forensic medicine. However, its utility in PMI estimation remains limited. Hanging cases have specific features of livor mortis that affect PMCT findings related to PMI. Recent artificial intelligence (AI) can evaluate images without bias, and may be useful for exploring which imaging findings could contribute to PMI estimation. This study aims to evaluate whether AI-based PMCT analysis can support PMI estimation in hanging cases.

Estimation of the postmortem interval (PMI) is important in forensic practice. In recent years, the use of postmortem CT (PMCT) has been increasing in the field of forensic medicine. However, its utility in PMI estimation remains limited. Hanging cases have specific features of livor mortis that affect PMCT findings related to PMI. Recent artificial intelligence (AI) can evaluate images without bias, and may be useful for exploring which imaging findings could contribute to PMI estimation. This study aims to evaluate whether AI-based PMCT analysis can support PMI estimation in hanging cases.

Methods

A retrospective review was conducted on all fatal hanging cases examined by PMCT and autopsy at the Department of Legal Medicine, Chiba University, over a ten-year period (2016–2025). Exclusion criteria included cases with a total body score (TBS) of the trunk ≥ 4, undocumented last-seen or discovery times, and individuals alive at discovery who achieved return of spontaneous circulation (ROSC).

The AI model used was DINOv3, which is a vision transformer based neural network architecture that had been pre-trained on general images, with chest-focused volume of interest (VOI) analysis. The time interval between last-seen alive and discovery was dichotomized as ≤ 0.5 day and > 0.5 day, using a single threshold to separate early and later discovery.

A retrospective review was conducted on all fatal hanging cases examined by PMCT and autopsy at the Department of Legal Medicine, Chiba University, over a ten-year period (2016–2025). Exclusion criteria included cases with a total body score (TBS) of the trunk ≥ 4, undocumented last-seen or discovery times, and individuals alive at discovery who achieved return of spontaneous circulation (ROSC).

The AI model used was DINOv3, which is a vision transformer based neural network architecture that had been pre-trained on general images, with chest-focused volume of interest (VOI) analysis. The time interval between last-seen alive and discovery was dichotomized as ≤ 0.5 day and > 0.5 day, using a single threshold to separate early and later discovery.

Results

A total of consecutive 145 fatal hanging cases were identified, of which 111 met the inclusion criteria. AI classification showed clear separation between the ≤ 0.5-day and >0.5-day groups in violin distribution plots. The model achieved area under the receiver operative characteristic curve (AUC) of 0.938.

A total of consecutive 145 fatal hanging cases were identified, of which 111 met the inclusion criteria. AI classification showed clear separation between the ≤ 0.5-day and >0.5-day groups in violin distribution plots. The model achieved area under the receiver operative characteristic curve (AUC) of 0.938.

Conclusion

AI can detect time-related differences in PMCT lung patterns between early and later discovery groups in fatal hanging and may serve as a supportive tool in PMI estimation.

AI can detect time-related differences in PMCT lung patterns between early and later discovery groups in fatal hanging and may serve as a supportive tool in PMI estimation.