Oral Cancer Recognition on Photographic Images Via Deep Learning Semantic Segmentation

Medical image segmentation is an important task supporting diagnosis and screening systems in several medical areas including oral cancer recognition. This paper explores the effectiveness of different deep learning (DL) architectures, including U-Net, LinkNet, PAN, and FPN for oral cavity lesion segmentation. Furthermore, we propose an ensemble model incorporating several decision fusion strategies to aggregate individual predictions, to improve the individual model performance. Our study employs a dataset acquired and manually labeled by the clinical subgroup of our team. On this dataset, we address two distinct segmentation problems: binary semantic segmentation to differentiate healthy tissue from diseased regions and multiclass semantic segmentation to identify three oral pathologies: aphthous, traumatic, and neoplastic lesions. We study the ensemble model's effectiveness in improving segmentation accuracy by combining different DL architectures' strengths. The results demonstrate that the ensemble strategy is highly effective for binary semantic segmentation, achieving a Dice score of 76.5%; while, for the multi-class problem of differentiating between multiple diseases, improvements are present but less marked.