Automatic measuring of coronary atherosclerosis from medicolegal autopsy photographs based on deep learning...

Area of Science:

• Forensic Medicine
• Cardiovascular Pathology
• Artificial Intelligence in Medicine

Background:

• Atherosclerotic cardiovascular disease diagnosis is crucial in forensic medicine due to its link with sudden death.
• Current visual assessment of coronary atherosclerosis in South Korean autopsies by forensic pathologists may lack diagnostic accuracy.
• Objective assessment of coronary atherosclerosis severity is needed for improved forensic pathology outcomes.

Purpose of the Study:

• To develop a deep learning algorithm for rapid and precise assessment of coronary atherosclerosis from autopsy images.
• To evaluate the performance of the developed algorithm in predicting atherosclerosis severity.
• To identify factors influencing the algorithm's predictions.

Main Methods:

• Retrospective analysis of 3,717 digital photographs from 1,920 forensic autopsies.
• Development of a deep learning algorithm using images of the left anterior descending and right coronary arteries.
• Validation of the algorithm's performance using agreement metrics and F1-scores on a test set.

Main Results:

• The deep learning algorithm showed high agreement (0.988) and absolute agreement (0.986) with ground truth values.
• The model achieved a strong overall weighted F1-score of 0.904.
• Class-wise F1-scores varied, with the lowest performance for moderate atherosclerosis (0.785). Calcification was a significant factor, unlike decomposition, stent implantation, or thrombi.

Conclusions:

• The developed deep learning algorithm shows significant potential as a practical tool for assessing coronary atherosclerosis in autopsy photographs.
• The AI model offers improved accuracy and speed compared to traditional visual assessments.
• Further research is warranted to enhance the model's performance, particularly for moderate atherosclerosis grades.