Abstract
Accurate diagnosis of oral lesions, early indicators of oral cancer, is a complex clinical challenge. Recent advances in deep learning have demonstrated potential in supporting clinical decisions. This paper introduces a deep learning model for classifying oral lesions, focusing on accuracy, interpretability, and reducing dataset bias. The model integrates three components: (i) a Classification Stream, utilizing a CNN to categorize images into 16 lesion types (baseline model), (ii) a Guidance Stream, which aligns class activation maps with clinically relevant areas using ground truth segmentation masks (GAIN model), and (iii) an Anatomical Site Prediction Stream, improving interpretability by predicting lesion location (GAIN+ASP model). The development dataset comprised 2765 intra-oral digital images of 16 lesion types from 1079 patients seen at an oral pathology clinic between 1999 and 2021. The GAIN model demonstrated a 7.2% relative improvement in accuracy over the baseline for 16-class classification, with superior class-specific balanced accuracy and AUC scores. Additionally, the GAIN model enhanced lesion localization and improved the alignment between attention maps and ground truth. The proposed models also exhibited greater robustness against dataset bias, as shown in ablation studies.