Semi-supervised labelling of the femur in a whole-body post-mortem CT database using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Forensic Science

Background:

• Accurate documentation of medical implants is crucial in forensic and autopsy cases.
• Current methods for identifying implants in post-mortem computed tomography (PMCT) can be time-consuming and may miss details not in reports.

Purpose of the Study:

• To develop and validate a deep learning pipeline for localizing and classifying femur implants in whole-body PMCT scans.
• To establish a proof-of-principle for automated, image-based labeling of post-mortem findings.

Main Methods:

• A deep learning pipeline integrating residual networks and an autoencoder was developed.
• The pipeline was trained and tested on 450 full-body PMCT scans.
• Localization accuracy was assessed using Dice scores and mean absolute errors; classification accuracy was evaluated for different implant types.

Main Results:

• Localization achieved high Dice scores (0.99 axial, 0.96 coronal, 0.98 sagittal) and low mean absolute errors (3.2-7.1 mm).
• Classification accuracy exceeded 97%, with high recall for 'without-implant' (1.00) and hip replacements (1.00), and lower recall for knee replacements (0.82) and nails (0.65).
• Implant orientation and extremity positioning were significant factors in localization accuracy.

Conclusions:

• The developed deep learning pipeline offers a robust, semi-automatic approach for implant detection and classification in PMCT.
• This method provides a generalized framework for image-based labeling, potentially improving the efficiency and completeness of post-mortem analyses.
• Further refinement may enhance recall for specific implant types, increasing its utility in forensic radiology.