Histopathological image analysis and enhanced diagnostic accuracy explainability for oral cancer detection

Deep learning (DL) has transformed medical imaging, particularly in the realm of oral cancer (OC) diagnosis using histopathological images. Timely detection of OC is es-sential for enhancing precision medicine and saving lives. However, incorrect diagnosis may impede effective treatment. In this study, we have proposed a DL model for OC classification, enhanced diagnosis decision-making and interpretability. We achieve this by starting with colour normalisation of histopathology images using the Vaha-dane 3-stain parameter normalisation and watershed segmentation method, followed by tiling and augmentation. Key features are selected using the weighted Fisher score (WFS) to address class imbalance. The U-Net classifier has been improved by using feature-based inputs instead of full images, reducing computational complexity and training time. The integration of Vahadane normalisation for consistent preprocessing across samples, WFS, and explainable artificial intelligence (XAI) addresses critical challenges in histopathological image analysis. The proposed model surpasses exist-ing approaches with a classification accuracy of 99.54%, and it outperforms Dense- Net201 and VGG10 in precision and reliability. The efficiency in handling imbal-anced datasets and explainability features make it suitable for early precise OC detec-tion, which can reduce diagnostic errors and enhance treatment outcomes.