Artificial intelligence tools for clinical decision support in dermatology
File(s)
Author(s)
Yilmaz, Abdurrahim
Type
Thesis or dissertation
Abstract
Dermatology is inherently multimodal, requiring the integration of clinical photographs, dermoscopy, microscopy, textual patient information, and longitudinal follow-up. Artificial intelligence (AI) has considerable potential to support clinical decision-making and enhance workflow efficiency. Recent advances in language models further extend these capabilities by enabling the joint interpretation of textual and visual clinical information and supporting more context-aware reasoning. However, the development and deployment of clinically useful dermatology AI remain constrained by fragmented datasets, scarce image-text supervision, privacy concerns, and the high computational demands of state-of-the-art models.
This thesis develops a data-centric and resource-efficient AI framework for dermatology, spanning dataset construction, model development, benchmarking, and privacy-preserving deployment. First, it establishes structured data foundations through DERM12345, a multisource dermoscopic dataset comprising 12,345 images organized into 40 clinically meaningful subclasses, and DermaSynth, a synthetic multimodal dermatology dataset generated from open-access datasets, PubMed-derived resources, and instruction-following strategies to provide scalable image-text supervision for dermatology vision-language models (VLMs). Together, these resources support fine-grained skin lesion analysis and multimodal learning. Second, it investigates a range of AI methods for dermatological applications, beginning with convolutional neural network (CNN) baselines on DERM12345 and progressing to artifact-aware object detection for fungal hyphae localization in potassium hydroxide (KOH) microscopy, local language models for longitudinal pemphigus follow-up, and DermatoLlama, a dermatology-specific vision-language model designed for report generation and multimodal reasoning under resource constraints. Third, it introduces evaluation infrastructure, including AnnotatorMed, a locally deployable hierarchical annotation tool; DermaBench, a clinician-annotated visual question-answering (VQA) benchmark integrating clinical and dermoscopic images; and hierarchical foundation-model benchmarking on DERM12345.
Together, these contributions define an end-to-end pathway for explainable, multimodal, resource-efficient, and privacy-preserving dermatology AI. The central argument of this thesis is that clinically credible dermatology AI must be not only accurate but also interpretable, evaluable, data-efficient, and compatible with real-world healthcare constraints.
This thesis develops a data-centric and resource-efficient AI framework for dermatology, spanning dataset construction, model development, benchmarking, and privacy-preserving deployment. First, it establishes structured data foundations through DERM12345, a multisource dermoscopic dataset comprising 12,345 images organized into 40 clinically meaningful subclasses, and DermaSynth, a synthetic multimodal dermatology dataset generated from open-access datasets, PubMed-derived resources, and instruction-following strategies to provide scalable image-text supervision for dermatology vision-language models (VLMs). Together, these resources support fine-grained skin lesion analysis and multimodal learning. Second, it investigates a range of AI methods for dermatological applications, beginning with convolutional neural network (CNN) baselines on DERM12345 and progressing to artifact-aware object detection for fungal hyphae localization in potassium hydroxide (KOH) microscopy, local language models for longitudinal pemphigus follow-up, and DermatoLlama, a dermatology-specific vision-language model designed for report generation and multimodal reasoning under resource constraints. Third, it introduces evaluation infrastructure, including AnnotatorMed, a locally deployable hierarchical annotation tool; DermaBench, a clinician-annotated visual question-answering (VQA) benchmark integrating clinical and dermoscopic images; and hierarchical foundation-model benchmarking on DERM12345.
Together, these contributions define an end-to-end pathway for explainable, multimodal, resource-efficient, and privacy-preserving dermatology AI. The central argument of this thesis is that clinically credible dermatology AI must be not only accurate but also interpretable, evaluable, data-efficient, and compatible with real-world healthcare constraints.
Date Issued
2026-06-22
Date Awarded
2026-09-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Temelkuran, Burak
Posma, Joram Matthias
Takats, Zoltan
Sponsor
Imperial College London.
Publisher Department
Department of Metabolism, Digestion and Reproduction
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)