AI-MODEL FOR THE DETECTION OF RARE SKIN DISEASE AND CANCER

Abstract

Early detection of rare skin diseases (RSD) and cancers is crucial for improving patient outcomes and addressing healthcare disparities. This thesis explores the transformative role of artificial intelligence (AI) in dermatology through a comprehensive scoping review and the development of two novel AI models. The review synthesizes 68 empirical studies and reveals a strong focus on diagnostic support using unimodal imaging data (e.g., dermoscopy, histopathology, immunofluorescence), while multimodal integration, prognostic modeling, and treatment planning remain underexplored. Key challenges identified include limited datasets, class imbalance, poor generalizability due to lack of external validation, and insufficient fairness evaluations—particularly for underrepresented populations. Only ~10% of studies apply multimodal fusion, and fewer than 2% use integrated or stacked model architecture. Promising techniques such as federated learning, few-shot learning, and attention mechanisms remain underutilized yet offer significant potential. In direct response to these gaps, two AI frameworks were developed to demonstrate practical, targeted solutions. EBAnet, based on EfficientNet and Grad-CAM, was trained on direct immunofluorescence (DIF) images for early detection of Epidermolysis Bullosa Acquisita (EBA), a disease notably underrepresented in the literature. It achieved 96.7% accuracy and an AUC of 0.994, with Grad-CAM offering interpretable visualizations for clinical insight. The second model, a stacked ensemble integrating CNNs, Swin/ViT transformers, and machine learning classifiers (e.g., XGBoost, TabNet), was applied to the ISIC 2024 dataset for rare skin cancer detection and achieved an AUC of 0.90067. These models were designed to reflect key recommendations from the review—emphasizing multimodal integration, robust evaluation, and interpretability. Both aim to bridge the translational gap between research and real-world clinical deployment. This thesis contributes practical AI tools and strategic insights for advancing precision dermatology. It underscores the need for standardized evaluation, fairness-aware design, and real-world validation to ensure equitable AI solutions for rare skin conditions. Future directions include integrating unstructured data such as clinical notes and genomics, expanding multi-institutional datasets, and aligning with regulatory pathways. Collectively, the work lays a strong foundation for next-generation AI systems in rare dermatology, moving toward personalized and inclusive patient care.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering