# Efficient and explainable histopathology for cancer detection using dual-teacher distillation and integrated gradients

## TL;DR

This paper introduces a fast and accurate AI model for detecting gastric cancer in histopathology slides, using a lightweight student network trained by two teacher models and providing interpretable results.

## Contribution

A novel dual-teacher knowledge distillation framework with integrated gradients for explainable gastric cancer classification.

## Key findings

- The MobileNet-V2 student model achieved 95.78% to 98.33% accuracy on multi-resolution patches.
- The student model was over thirty times smaller and twice as fast as the teacher models.
- Integrated Gradients highlighted meaningful regions like nuclei clusters and gland boundaries for model interpretability.

## Abstract

Gastric cancer remains one of the most common malignancies worldwide. The timely and accurate histopathological diagnosis plays a critical role in effective treatment. Manual interpretation of histology slides is time consuming and requires considerable expertise. To address these challenges, this study introduces a two-teacher one-student (2T–1S) knowledge distillation framework for gastric cancer classification using the GasHisSDB dataset. The framework leverages DenseNet-121 and ResNet-50 as teacher networks to guide a lightweight MobileNet-V2 student. This approach provided high accuracy with significantly reduced computational cost. Experiments on multi-resolution patches (80 × 80, 120 × 120, and 160 × 160) show that the MobileNet-V2 student achieved accuracies of 95.78%, 97.46%, and 98.33%, respectively. Also, the teacher model DenseNet-121 achieved the accuracies of 96.44%, 98.75% and 98.19% and the ResNet-50 teacher reached 96.63%, 97.87% and 98.31% respectively. In addition, the student network was more than thirty times smaller and nearly twice as fast during inference. This fast light-weight model is well-suited for real-time inference on resource-constrained devices. Integrated Gradients were applied to explain the model was paying attention to actual features and focus on meaningful regions like nuclei clusters and gland boundaries. Compared with many existing techniques this framework act as balance trade-off between accuracy, speed and interpretability. This balance positions the framework as a viable tool for digital pathology workflows and further refinement could extend its utility to clinical decision support.