An Enhanced U-Net Model for Precise Oral Epithelial Layer Segmentation Using Patch-Based Training

Abstract

The oral epithelial layer is crucial for detecting oral dysplasia and cancer from histopathology images. Accurate segmentation of the oral epithelial layer in biopsy slide images is essential for early detection and effective treatment planning of conditions like Oral Epithelial Dysplasia, where abnormal changes increase the risk of oral cancer. This study investigates using a Deep Learning model to precisely identify and segment areas of the Oral Epithelial Layer in biopsy images of the oral cavity, aiming to enhance early diagnosis and treatment strategies. The study is conducted with an indigenously collected and benchmarked dataset of 300 histopathology images of the oral cavity, representing 64 patients. We propose a Deep Learning-based modified U-Net model for segmenting oral cavity histopathology images. Various patch sizes and batch size combinations were tested and implemented for comparison. The performance of the optimal patch and batch size combination is further compared with relevant state-of-the-art models. The modified U-Net model utilizing the patch generation technique demonstrated superior performance in oral cavity epithelium segmentation, achieving an IoU of 98.06, precision of 99.66, recall of 99.13, and F1-score of 99.00. Our research underscores the efficacy of deep learning-based segmentation with the patch generation technique in improving oral health diagnostics, outperforming several state-of-the-art models in segmenting the epithelial layer. This research enhances segmentation, a key step in Computer-Aided Diagnosis systems, ensuring accurate analysis, efficient processing, and reliable medical image interpretation for improved patient outcomes.