A Deep Feature Driven Expert System to Estimate the Postmortem Interval From Corneal Opacity Development

Abstract

Postmortem interval (PMI) estimation remains an unresolved challenge in forensic science, necessitating practical, reliable and more accurate tools. This study aimed to develop a quantitative PMI estimation tool that effectively meets these needs. Focusing on the postmortem opacity development of the eye as a key marker for determining time since death, we propose an artificial intelligence-based clinical PMI prediction system utilising computer vision, deep learning and machine learning methods. The AlexNet algorithm was utilised to extract deep features from the postmortem eye images. Extracted features were then processed by machine learning algorithms. For feature selection, Lasso and Relief techniques were employed, while SVM and KNN were applied for classifications. The results were validated using the leave-one-subject-out method. The system was tested across different postmortem ranges, providing multi-label predictions. The performance was evaluated using various metrics. The deep features exhibited effective performance in grading postmortem opacity development, achieving state-of-the-art results. The accuracy scores were 0.96 and 0.97 for 3-h intervals (i.e., 5-class) and 5-h intervals (i.e., 3-class) experiments, respectively. The experimental results indicate that the proposed system represents a promising tool for PMI estimation.