Preserving Spatial and Hierarchical Lesion Structures in Automated Skin Cancer Diagnosis through DermaCap and CCSA

Abstract

Early and reliable detection of skin cancer remains a global clinical priority, particularly for rare lesion categories that are often underrepresented due to limited data and severe class imbalance. To address these challenges and the limitations of traditional Convolutional Neural Networks (CNNs) in preserving spatial relationships, this study proposes DermaCap, a hybrid CNN–Capsule Network architecture designed to capture lesion pose, orientation, and hierarchical structural features. The framework follows a rare-class-first design philosophy, optimizing the model architecture, dataset curation, and training strategy to improve sensitivity toward diagnostically challenging lesions. Additionally, Clinically Constrained Structural Augmentation (CCSA) is introduced to enhance dataset diversity while preserving dermatologically meaningful morphology and color realism. The proposed model was evaluated using a carefully selected six-class subset of the HAM10000 dermatology database, which contains a total of 3,310 dermatoscopic images representing multiple diagnostic categories. The evaluation resulted in an approximate accuracy of 91%. The precision, recall, and F1-score show consistent values across both majority and minority classes. Therefore, the proposed model shows stable and consistent diagnostic performance for dermatoscopic image classification. Overall, DermaCap presents a clinically aligned, structure-aware approach for more trustworthy and scalable AI-assisted skin lesion diagnosis.