Artificial intelligence in forensic pathology: Multi-organ postmortem pathomics for estimating postmortem interval

Abstract

Background

Accurate estimation of the postmortem interval is crucial in forensic investigations. Pathomics presents a promising advancement by leveraging whole-slide images as a novel data modality for the diagnosis and prognosis of diseases in clinical situations. The extended application of this technology in forensic postmortem image analysis is expected to give rise to postmortem pathomics as an important subfield.

Objective

This study aimed to develop a three-level hierarchical strategy using pathomics to analyze postmortem histological images data, develop multi-organ integrated model for the postmortem interval estimation, and lay the foundation for postmortem pathomics.

Methods

Twelve Bama miniature pigs were euthanized, and liver, kidney, and skeletal muscle tissues were collected at 6, 24, 48, and 96 h postmortem. Hematoxylin and eosin stained whole slide images were divided into 512 × 512 pixel patches. Low-quality patches were excluded using Otsu thresholding, and color normalization was applied using the Vahadane algorithm to minimize staining variability. Deep learning models were trained on patch-level data using transfer learning and evaluated for interpretability with Grad-CAM. Slide-level predictions were obtained via organ-specific deep feature aggregation and machine learning models, while a multi-organ integrated model was developed using a stacking ensemble combining above machine learning models and a logistic regression. Four additional pigs were introduced for external validation at the multi-organ integrated individual-level.

Results

DenseNet121 demonstrated superior performance for liver and kidney, while VGG16 performed best for skeletal muscle tissue. These models were designated as postmortem-liver-net, postmortem-kidney-net, and postmortem-muscle-net, respectively, and employed to extract pathomics features from images. Slide-level models trained on these features vectors achieved accuracies of 81.25% (liver), 87.5% (kidney), and 62.5% (muscle). A stacking model integrating probability output of these three slide-level models achieved internal and external test accuracies at multi-organ integrated individual-level of 93.75% and 87.5%, respectively.

Conclusion

This study demonstrated the potential of pathomics and deep learning for postmortem interval estimation. The proposed three-level framework effectively integrated multi-organ features, introducing whole-slide images as a novel modality and offering an innovative strategy for postmortem interval estimation.