Oral cancer (OC) is a prevalent malignancy in India, ranking sixth nationally and thirteenth globally. The ICMR National Cancer Repository Program forecasts 15.7 lakh cancer cases by 2025, up from 14.6 lakh in 2014. India’s diverse cancer occurrence patterns pose significant challenges to prevention and treatment. Regular health checks and increased risk factor awareness are crucial for improving OC survival rates. In this study, we explore the field of oral epithelial dysplasia pathology by using deep learning (DL)-based models to segment 900 carefully curated images of the histopathology of oral epithelial dysplasia at a 100x magnification from the biopsy slides. We use a two-pronged training strategy that includes the Vanilla U-Net model and a fine-tuned U-Net variation. Our results show significant differences in performance, with the optimized U-Net model performing better. To be more precise, the modified U-Net model produces the following results: Intersection over Union (IoU): 95.24%; Precision: 98.77%; Recall: 98.19%; and F1_Score: 97.55%. Additional evaluations utilizing additional native histopathology pictures of OC validate the resilience of our training models, resulting in the effective segmentation of epithelial layers. This study adds to our understanding of the pathophysiology of OC and highlights the potential of DL methods for accurate diagnosis and treatment.