Multimodal Integration for Translational AI in Dermatology

Abstract

Dermatologic imaging plays a central role in diagnosis, disease monitoring, and clinical decision-making. The increasing availability of large-scale dermatologic imaging data and AI advances introduce challenges in data management, multimodal integration, and evaluation of emerging generative AI methods in dermatology. This dissertation addresses these challenges by extending structured frameworks from neuroimaging, a well-established domain for AI and medical image analysis, to enable multimodal integration for translational AI in dermatology. First, we characterize developmental brain signatures across the superficial and deep white matter by adapting multimodal MRI approaches to map short association fibers in relation to long-range white matter tracts, establishing our conceptual foundation for multimodal integration. Second, we develop the Dermatology Imaging Data Structure (DermIDS), an AI-ready structure inspired by BIDS in neuroimaging, to integrate and manage heterogeneous dermatologic imaging data in non-PHI research systems. Curating over a million skin images across clinical and computer vision datasets, we demonstrate that DermIDS supports multimodal integration at scale and identifies inconsistencies in how images are characterized, supporting quality control and harmonization. Third, we translate and develop generative AI models to evaluate emerging privacy risks and multimodal representations for dermatologic image analysis. We empirically assess the effectiveness of conventional facial masking strategies and provide quantitative metrics of reidentification risk in the context of generative AI. We also systematically evaluate AI-generated representations as a potential strategy to enrich dermatologic image datasets, characterizing generalization and failure modes across clinical images. Collectively, this dissertation establishes a framework for structured data management, multimodal integration, and evaluation of emerging generative AI technologies in dermatologic image analysis.