Abstract
In forensic pathology, identifying the origin of tissue samples is a critical yet challenging task due to postmortem changes such as autolysis and putrefaction. Traditional clinical AI tools are often ill-equipped for the structural distortions found in forensic samples. We present FPath, a novel deep learning framework that integrates a hybrid CNN-Transformer backbone with self-supervised contrastive learning and a context-aware Multiple Instance Learning (MIL) network. FPath achieves state-of-the-art performance, reaching 98.31% accuracy on rat tissues and 90.49% on real-world human forensic samples, demonstrating robust cross-species generalization and high diagnostic potential.
Introduction
Identifying the origin of a tissue sample is a fundamental step in forensic pathology for determining cause and time of death. Unlike clinical pathology, where samples are typically fresh, forensic samples undergo significant postmortem degradation. Processes like putrefaction and autolysis destroy cellular structures, causing different organs to appear confusingly similar.
Existing AI tools, primarily optimized for clinical tasks like cancer detection, struggle with the "noisy" visual data inherent in forensic contexts. There is a pressing need for systems capable of learning fine-grained, discriminative representations that remain robust despite postmortem distortions.
Methodology
The FPath framework employs a two-step architecture designed to capture both local cellular details and broader structural context.
1. Self-Supervised Double-Tier Backbone
To address the loss of typical cellular structures, FPath uses a hybrid backbone:
- Local Branch (ResNet50): Captures fine textures and local cellular details.
- Global Branch (Swin Transformer): Captures long-range spatial relationships and organ-level structures.
The training utilizes a self-supervised contrastive learning strategy with a multi-part loss function. This includes a global cosine loss, a "parts" loss to encourage local region similarity, and variance/covariance regularization (inspired by VICReg) to prevent feature collapse.
2. Context-Aware Multiple Instance Learning (MIL)
Whole-slide images are processed as collections of patches (instances). FPath employs a multi-head self-attention mechanism with relative positional embeddings, allowing patches to interact and refine features based on neighboring context. An adaptive pooling operation then filters uninformative patches, highlighting discriminative features for final classification.
Experimental Setup and Datasets
The study leverages two distinct datasets:
- Rat Postmortem Histopathology Dataset: Comprising 19,607 images from 90 rats across seven organs (brain, heart, kidney, liver, lung, pancreas, spleen) at multiple magnifications.
- Human Forensic Histopathology Dataset: Consisting of 3,378 real-world images from 32 human decedents. This dataset provides a rigorous test for cross-domain generalization.
Results and Discussion
FPath demonstrated superior performance compared to several state-of-the-art self-supervised and MIL methods, including TransPath, RetCCL, and MOCOV3.
Key results include:
- High Accuracy: FPath achieved near-perfect accuracy (98.31%) on the rat dataset and maintained high performance (90.49%) on the more challenging human dataset.
- Cross-Species Generalization: The model's ability to generalize from rat to human tissues suggests that the fundamental histopathological patterns of decay are somewhat conserved across species, or that the learned features are sufficiently robust to bridge the gap.
- Explainability: Visualization using LayerCAM confirmed that FPath focuses on biologically relevant structures, such as kidney glomeruli or brain meningeal structures, even when degraded.
Conclusion and Future Work
FPath represents a significant advancement in automated forensic pathology. By combining hybrid architectures with self-supervised learning, it effectively handles the complexities of postmortem tissue recognition. Future research will focus on expanding the human dataset and extending the framework to diagnostic tasks, such as estimating the postmortem interval or identifying specific pathological changes related to the cause of death.
Summary updated