RiFNet: Automated rib fracture detection in postmortem computed tomography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Forensic Radiology

Background:

• Medical imaging is crucial for diagnostics, but manual analysis can cause delays.
• Automated analysis of medical images, particularly postmortem computed tomography (PMCT), is needed to overcome expert shortages.
• Rib fractures are common findings in PMCT, requiring accurate detection for forensic and clinical purposes.

Purpose of the Study:

• To develop and evaluate a custom convolutional neural network (RiFNet) for detecting rib fractures in PMCT scans.
• To compare the performance of RiFNet against established transfer learning models (ResNet50 V2, Inception V3).
• To assess the suitability of transfer learning for rib fracture classification in PMCT data.

Main Methods:

• A retrospective cohort of 195 postmortem cases with rib fractures was analyzed.
• PMCT data were processed using Syngo.via software, unfolding the rib cage into single images.
• A custom CNN (RiFNet) and two pre-trained CNNs (ResNet50 V2, Inception V3) were trained and validated on 585 images.

Main Results:

• RiFNet achieved a high average F1 score of 0.91 ± 0.04.
• Transfer learning models achieved lower F1 scores: Inception V3 (0.64) and ResNet50 V2 (0.61).
• RiFNet demonstrated superior efficiency and accuracy in detecting rib fractures on PMCT compared to transfer learning.

Conclusions:

• RiFNet is an effective deep learning tool for automated rib fracture detection in postmortem CT.
• Standard transfer learning approaches may not be optimal for the specific characteristics of PMCT data.
• AI-driven analysis holds significant potential to improve the efficiency of forensic radiology workflows.