INTEGRATING HARALICK TEXTURES AND SHAP VALUES FOR INTERPRETABLE PATCH-BASED ORAL CANCER CLASSIFICATION

Accurate detection and classification of oral cancer is still a challenge due to the complex morphology of oral tissues. Processing high-resolution histopathological images using deep learning models is computationally complex and leads to the proposal of patch-based methods. However, traditional patch selection techniques, which rely on nuclear density or random sampling, often miss diagnostically important regions, resulting in a minimum performance. Furthermore, most existing methods lack explainability in the patch selection process, limiting clinical interpretability. To address these challenges, we propose an explainable patch selection algorithm that introduces a novel metric, termed the patch score, designed to identify the most informative regions in oral histopathology images. This enhances explainability, a critical requirement for clinical trust in automated systems. Haralick texture features extracted from training images are used to train a neural network classifier. Once the model is trained, we apply SHapley Additive exPlanations (SHAP) to compute the importance values for each Haralick feature. These SHAP values reveal how each individual feature influences the model’s predictions, thereby improving the transparency and interpretability of the model’s decision-making process. During testing, each image is first preprocessed and then divided into 50% overlapping patches. Haralick features are extracted from each patch, and their similarity to the global SHAP value vector (obtained from the full image) is used to score and rank the patches for further feature extraction. The patch score is calculated using the SHAP values of the test image and the features of each patch, enabling the selection of the most informative regions. The selected patches are then processed through a Dual Path Network with dilated convolution (DPN-92D) to extract deep features for classification. Our method, evaluated on an oral histopathology dataset with known cancer labels, achieved a classification accuracy of 97.88% and demonstrated improved computational efficiency. Overall, the proposed method offers a robust, interpretable, and efficient Computer-Aided Diagnosis (CAD) solution to assist pathologists in the early and accurate diagnosis of oral cancer.