PMI estimation with cross-species transfer learning and visual information generated by pathomics foundation model.

Area of Science:

• Forensic Science
• Artificial Intelligence
• Pathology

Background:

• Estimating postmortem interval (PMI) is crucial in forensic science.
• Deep learning in pathology image analysis shows promise for PMI estimation.
• Existing methods struggle with animal-to-human sample translation and model interpretability.

Purpose of the Study:

• To develop a robust postmortem interval estimation framework.
• To address challenges in cross-species data translation and model interpretability in forensic pathology.
• To create an interpretable AI tool for forensic analysis.

Main Methods:

• A pathomics foundation model with a two-stage cross-species transfer learning strategy was proposed.
• The model was fine-tuned on porcine liver whole-slide images (WSIs) followed by human data.
• Visualization techniques including probability maps and classification proportion histograms were used for interpretability.

Main Results:

• The Vision Transformer-based UNI model achieved 91.63% accuracy on porcine data.
• Transfer learning with limited human data improved accuracy to 78.95%, a >50% increase.
• The visualization framework enhanced model interpretability and traceability.

Conclusions:

• Combining animal data priors with fine-tuning on minimal human data enables cross-species PMI estimation.
• The framework effectively addresses data scarcity and enhances AI model transparency in forensic pathology.
• A practical and interpretable AI-based tool for forensic pathology was developed.