Gray-Scale Extraction of Bone Features from Chest Radiographs Based on Deep Learning Technique for Personal...

Area of Science:

• Forensic Radiology
• Medical Imaging Analysis
• Artificial Intelligence in Pathology

Background:

• Post-mortem (PM) imaging aids individual identification by comparing with ante-mortem (AM) scans.
• Bone structures in radiographic images are crucial for identification but obscured by soft tissue decomposition in PM images.
• Accurate bone image extraction is vital for reliable forensic identification.

Purpose of the Study:

• To evaluate the effectiveness of U-Net for extracting bone images from 2D X-ray data.
• To assess U-Net's performance on both simulated PM and real AM radiographic images.
• To determine the utility of extracted bone images for personal identification in forensic contexts.

Main Methods:

• U-Net model was trained using pseudo 2D X-ray images derived from PM computed tomography (CT) data.
• Two image types were generated: projections of all tissues and projections of bones only.
• Performance metrics included Intersection over Union, Dice coefficient, and area under the receiver operating characteristic curve.

Main Results:

• U-Net successfully extracted bone images visually and accurately from both AM and PM X-ray datasets.
• The method demonstrated robust performance on real 2D AM chest radiographs.
• Quantitative evaluations confirmed the high accuracy of the bone extraction process.

Conclusions:

• U-Net is an effective tool for bone image extraction from 2D X-rays, even with post-mortem decomposition.
• Extracted bone images provide valuable information for personal identification in forensic pathology.
• This AI-driven approach enhances the reliability of forensic identification using radiographic imaging.