Scholarlink

Background <p>The determination of the postmortem interval (PMI) is essential for forensic investigations. Postmortem computed tomography (PMCT) offers a&#xa0;noninvasive method of visualising postmortem changes. However, analysing these images is time-consuming. Deep learning could automate this process and enable more objective predictions.</p> <p>This study investigates whether a&#xa0;three-dimensional convolutional neural network (3D-CNN) can classify the PMI using nonsegmented PMCT data.</p> Materials and methods <p>A&#xa0;total of 544&#xa0;PMCT scans were prospectively and serially obtained from 123&#xa0;deceased persons at various time intervals. The initial CT scan had a&#xa0;postmortem interval (PMI) of less than 6 h in each case. The scans were assigned to three PMI classes (0–6&#xa0;h, 9–24&#xa0;h, 36–96&#xa0;h). A&#xa0;modified 3D ResNet-18 architecture was used for modelling, which was tested with different resolutions (64 × 64, 128 × 128, 256 × 256 pixels) and slice depths (16, 64, 128 slices).</p> Results <p>The best model performance was achieved with 128 × 128 pixels and 128 slices. Lower resolutions (64 × 64) resulted in near-random predictions, while higher resolutions (256 × 256) showed no significant improvements. The classification of the medium PMI class was the most reliable, whereas the early PMI class performed poorly.</p> Conclusions <p>The findings indicate that 3D-CNNs can be useful for PMI estimation, but larger and standardized datasets are required. Focusing on specific anatomical regions (e.g., head or thorax) could enhance model accuracy by reducing irrelevant information from whole-body scans.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Einsatz eines 3D-Convolutional Neural Network zur Prädiktion des postmortalen Intervalls in der postmortalen Computertomographie

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月21日

查看全文
  • Julian Ivanov,
  • Benjamin Ondruschka,
  • Axel Heinemann

摘要

Background

The determination of the postmortem interval (PMI) is essential for forensic investigations. Postmortem computed tomography (PMCT) offers a noninvasive method of visualising postmortem changes. However, analysing these images is time-consuming. Deep learning could automate this process and enable more objective predictions.

This study investigates whether a three-dimensional convolutional neural network (3D-CNN) can classify the PMI using nonsegmented PMCT data.

Materials and methods

A total of 544 PMCT scans were prospectively and serially obtained from 123 deceased persons at various time intervals. The initial CT scan had a postmortem interval (PMI) of less than 6 h in each case. The scans were assigned to three PMI classes (0–6 h, 9–24 h, 36–96 h). A modified 3D ResNet-18 architecture was used for modelling, which was tested with different resolutions (64 × 64, 128 × 128, 256 × 256 pixels) and slice depths (16, 64, 128 slices).

Results

The best model performance was achieved with 128 × 128 pixels and 128 slices. Lower resolutions (64 × 64) resulted in near-random predictions, while higher resolutions (256 × 256) showed no significant improvements. The classification of the medium PMI class was the most reliable, whereas the early PMI class performed poorly.

Conclusions

The findings indicate that 3D-CNNs can be useful for PMI estimation, but larger and standardized datasets are required. Focusing on specific anatomical regions (e.g., head or thorax) could enhance model accuracy by reducing irrelevant information from whole-body scans.