The fragmentation problem in dermatology AI Existing AI tools for skin disease are trained on narrow single-modality, single-institution datasets that perform well on their specific benchmark but fail to generalize across imaging modalities, disease types, or clinical settings. A dermatologist uses total body photography, dermoscopy, clinical photography, and dermatopathology slides interchangeably - no prior AI model could operate across all four.
PanDerm architecture PanDerm is a multimodal vision foundation model pretrained using self-supervised learning on over 2 million real-world skin disease images from 11 institutions across 4 imaging modalities: total body photography tiles, dermatopathology slides, clinical photographs, and dermoscopy images. Self-supervised pretraining allows the model to learn rich visual representations without requiring manual disease labels for every image.
Performance summary Evaluated across 28 clinical benchmarks, PanDerm outperformed all task-specific AI models and clinician baselines. Key highlights include: outperforming clinicians by 10.2% in early-stage melanoma detection using longitudinal total body photography analysis, improving clinician accuracy by 11% when used as a decision-support tool on dermoscopy, and enhancing non-dermatologist differential diagnosis by 16.5% across 128 skin conditions.
Self-supervised learning strategy PanDerm was pretrained using masked image modeling - a self-supervised technique where portions of each image are masked and the model learns to reconstruct them. This forces the model to learn meaningful visual representations of skin disease pathology without requiring expensive disease labels for the pretraining corpus, enabling effective use of the full 2 million image dataset.
Multimodal pretraining dataset The 2 million image pretraining corpus spans diverse real-world clinical settings across 11 institutions in multiple countries. The four modalities - total body photography tiles, clinical photographs, dermoscopy images, and dermatopathology slides - correspond to distinct diagnostic workflows that clinicians use at different stages of patient evaluation.
Foundation model paradigm Following the paradigm established by large language models, PanDerm's representations are then adapted to specific downstream tasks using small labeled datasets via fine-tuning or linear probing. This separation of large-scale pretraining from task-specific adaptation means the foundation model can be deployed across dozens of clinical tasks without retraining from scratch.
Longitudinal total body photography analysis One of PanDerm's most clinically significant capabilities is longitudinal melanoma surveillance using total body photography. By comparing sequential TBP images from the same patient over time, PanDerm identifies changing lesions that may represent early melanoma - the exact clinical scenario where early intervention most dramatically improves prognosis.
10.2% improvement over clinicians In formal evaluation against dermatologist baselines, PanDerm outperformed clinicians by 10.2% in early-stage melanoma detection from longitudinal TBP analysis. This represents a genuine clinical advance, as early-stage melanomas are the most curable and the most likely to be missed on routine surveillance.
Dermoscopy decision support Beyond independent diagnosis, PanDerm was evaluated as an AI assistant for clinicians reading dermoscopy images. Dermatologists using PanDerm as a decision-support tool improved their accuracy by 11% compared to unaided assessment - demonstrating that the model adds genuine value even when paired with expert clinicians rather than replacing them.
28-benchmark evaluation PanDerm was systematically evaluated across 28 clinical benchmarks spanning melanoma detection, non-melanoma skin cancer classification, inflammatory dermatoses, rare diseases, and multi-class differential diagnosis tasks. Across these benchmarks, PanDerm achieved state-of-the-art performance compared to both specialist AI models and generalist vision models.
Non-dermatologist benefit The study specifically tested whether PanDerm could enhance the diagnostic capability of non-dermatologist clinicians - primary care physicians and emergency medicine providers who encounter skin disease but lack specialist training. Providing PanDerm-generated differential diagnoses improved non-dermatologist accuracy by 16.5% across 128 skin conditions, demonstrating potential to extend dermatologic expertise to non-specialist settings.
Cross-institutional generalization Unlike task-specific models that degrade sharply when applied to data from institutions not represented in training, PanDerm's diverse 11-institution pretraining corpus enables robust generalization. This is critical for real-world deployment where imaging hardware, preparation protocols, and patient demographics differ from any single training center.
Decision support rather than autonomous diagnosis The study frames PanDerm as a clinical decision support tool that augments rather than replaces dermatologists. The 11% accuracy improvement when used alongside clinicians - rather than simply measuring the model's standalone accuracy - is the metric most directly relevant to how AI will be deployed in actual practice.
Access equity implications The 16.5% non-dermatologist accuracy improvement is particularly significant for global health equity. Many low- and middle-income countries lack sufficient dermatology specialists; AI decision support that elevates non-specialist performance could extend early melanoma and skin cancer detection to geographies where specialist access is unavailable.
Multi-institution deployment considerations Real-world deployment requires regulatory clearance, liability frameworks, electronic health record integration, and workflow design. PanDerm's strong cross-institutional generalization reduces the calibration effort required when deploying to new sites, but prospective clinical trial data demonstrating patient outcome improvement will be required for regulatory approval.
Expanding beyond visual modalities The current PanDerm model integrates four visual imaging modalities, but dermatology diagnosis also incorporates patient history, genetic testing, reflectance confocal microscopy, and optical coherence tomography. A future multimodal foundation model incorporating clinical metadata alongside imaging could further improve diagnostic accuracy for ambiguous cases.
Prospective clinical trials All benchmarks reported are retrospective analyses. Demonstrating that PanDerm-assisted decision making actually improves patient outcomes - earlier melanoma stage at diagnosis, reduced unnecessary biopsies, or survival benefit - requires prospective randomized trials where AI assistance is randomized within clinical practices.
Rare disease and underrepresented populations Foundation models trained primarily on data from high-income, majority-population institutions may underperform on rare skin diseases and skin conditions in darker skin tones. Targeted data collection from underrepresented populations and rare disease registries is needed to ensure equitable performance and prevent AI-mediated diagnostic disparities.