Deep learning in dermatopathology: applications for skin disease diagnosis and classification

Abstract

Medical image segmentation is pivotal in disease diagnosis and treatment planning across various imaging modalities, including MRI, CT, ultrasound, X-ray, dermoscopy, and histopathology. This systematic literature review, conducted using the PRISMA framework, provides a comprehensive analysis of Deep Learning approaches applied to medical image segmentation, with a focus on dermato-pathology for skin disease diagnosis and classification. Transformer-based models have shown notable improvements over traditional CNN architectures, achieving up to 79.95% accuracy in multitask cancer detection tasks, surpassing CNN-based models that achieved 74.05%. In liver lesion segmentation using CT scans, attention-enhanced U-Net models achieved a 93.4% Dice Similarity Coefficient (DSC) for liver tissue and 77.8% for tumor segmentation. In dermoscopy, self-supervised transformer-based models like G2LL exceeded 80% accuracy, while U-Net-based models for skin lesion segmentation achieved up to 93.32% accuracy. Histopathology image analysis further demonstrated that models incorporating attention mechanisms, such as the PistoSeg framework, improved segmentation precision by up to 7.15% compared to conventional methods. Across various modalities, Deep Learning models consistently outperform traditional methods, with improvements ranging from 5 to 15% in accuracy and segmentation metrics. Despite challenges such as computational demands and the need for large annotated datasets, Deep Learning continues to revolutionize medical image segmentation, offering higher diagnostic precision and outlining future research directions to bridge existing gaps.