Multi-dimensional deep learning-based segmentation and volumetric assessment of sphenoid sinus fluid on postmortem CT...

Area of Science:

• Forensic Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Sphenoid sinus fluid is a key indicator in drowning cases.
• Manual assessment of this fluid on postmortem CT scans is time-consuming and subjective.
• There is a need for automated, quantitative methods in forensic analysis.

Purpose of the Study:

• To develop and evaluate deep learning models for automated segmentation and volumetric estimation of sphenoid sinus fluid.
• To assess the accuracy and reproducibility of these models compared to manual standards.
• To enable efficient, quantitative analysis of sphenoid sinus fluid in forensic practice.

Main Methods:

• Developed three U-Net-based deep learning models (2D, 2.5D, 3D) for sphenoid sinus fluid segmentation.
• Utilized postmortem CT images from 165 autopsy-confirmed drowning cases.
• Evaluated models against manual annotations for segmentation accuracy (Dice coefficient) and volumetric estimation (Spearman's correlation, mean absolute error).

Main Results:

• The 2D and 2.5D models achieved high mean Dice coefficients (0.866 and 0.869, respectively).
• Volumetric estimates showed strong correlations with reference standards (Spearman's ρ = 0.976-0.988) and no significant differences.
• The 2.5D model offered the best balance between segmentation and volume estimation accuracy.

Conclusions:

• Deep learning models can accurately automate sphenoid sinus fluid segmentation and quantification on PMCT images.
• This automated approach provides a feasible, reproducible, and efficient tool for forensic drowning investigations.
• Quantitative assessment of sphenoid sinus fluid is possible using PMCT images prior to autopsy.