Enhancing histopathological image analysis: An explainable vision transformer approach with comprehensive interpretation methods and evaluation of explanation quality

Deep learning models are increasingly reshaping medical imaging, with growing attention on ensuring transparency and trust in their decision-making processes. This study presents the Explainable Vision Transformer (XViT), a model specifically designed for histopathological image analysis. By incorporating advanced interpretability techniques, the XViT model addresses three core aspects: feature learning and classification, generating explainable outputs, and qualitatively evaluating these explanations. Three novel interpretability methods are introduced: attention-based, model-agnostic, and gradient-based, offering diverse perspectives on model behavior. The model's performance and generalizability were rigorously evaluated on two histopathological datasets: lung colon 25000 (LCS25000) with 96.2% accuracy across three classes and Kangbuk Samsung Hospital (KBSMC) with 88.6% accuracy across four classes. XViT provides actionable insights by highlighting diagnostically relevant regions in input images, significantly enhancing clinical trust and decision-making. The evaluation of its explainability methods through metrics like sensitivity, faithfulness, and complexity demonstrated that layer-wise relevance propagation for transformers outperforms standard techniques like local interpretable model-agnostic explanations (LIME) and attention visualization. This robust performance underscores the XViT model's potential to bridge the gap between AI accuracy and interpretability in medical imaging. Our findings emphasize the need for well-defined evaluation criteria when comparing interpretability methods and highlight the model's potential for integration into clinical workflows. This work represents a step forward in creating reliable and interpretable AI solutions, ensuring that the benefits of advanced deep learning models extend seamlessly into practical healthcare settings.