The Research Gap Previous studies of AI in dermatology reported that AI systems exceeded expert dermatologist accuracy, but these studies used curated benchmark datasets that did not reflect real-world clinical conditions. Key real-world challenges - including images from different institutions, disease categories not seen during training, and various image artifacts - were not systematically evaluated.
Challenge Design The 2019 ISIC Grand Challenge was designed to fill this gap by creating the largest public dermoscopy dataset (BCN20000) and evaluating AI performance under clinically realistic conditions. 64 teams submitted 129 state-of-the-art algorithm predictions on a test set of 8238 images, making this one of the largest AI skin cancer classification evaluations ever conducted.
Key Findings Summary The best performing algorithm achieved only 58.8% balanced accuracy on the BCN20000 realistic dataset, compared to 82.0% on the previously used HAM10000 benchmark. The dramatic performance drop reveals that AI systems developed on curated benchmarks substantially overestimate real-world clinical performance.
Clinical Safety Concern The most alarming finding was that 47.1% of images from disease categories not included in AI training were misclassified as malignant. If deployed clinically, this would generate substantial numbers of unnecessary biopsies, exposing patients to harm and generating healthcare costs without benefit.
Training Dataset Challenge participants received 25,331 training images from two institutions: HAM10000 from the Medical University of Vienna and the newly created BCN20000 from Hospital Clinic Barcelona. Images covered eight skin disease categories including nevi, melanoma, benign keratosis, dermatofibroma, basal cell carcinoma, squamous cell carcinoma, vascular lesions, and actinic keratosis. Images were collected between January 2000 and December 2018.
Test Dataset Design The 8238-image test set included images from both training institutions and from countries not represented in training data (Turkey, New Zealand, Sweden, and Argentina). Critically, the test set also included images of disease categories not present in training, grouped as a single 'not trained' (NT) category, testing whether algorithms could identify when they were operating outside their knowledge.
Ground Truth Standards Diagnostic labels were generated through histopathological review for all malignant and biopsied lesions, and through unanimous expert consensus (at least three board-certified dermatologists), digital monitoring, or confocal microscopy for unbiopsied benign lesions. This rigorous ground truth generation is necessary for meaningful performance evaluation.
Artifact Annotation An important innovation was manual annotation of image artifacts including ulceration, crust, pigmentation, hair, and pen markings. This enabled systematic analysis of how individual artifact types affect classification accuracy, providing specific guidance for algorithm improvement and deployment context limitations.
Dataset Distribution Shift The best performing algorithm achieved 82.0% balanced accuracy on HAM10000 (the 2018 benchmark) but only 58.8% on BCN20000, designed to better reflect real-world clinical conditions. This 23 percentage point performance drop quantifies how severely standard benchmarks overestimate real-world AI performance.
Geographic Generalization Failure Performance on images from Turkey, New Zealand, Sweden, and Argentina - countries not represented in training data - was lower than performance on images from the training institutions. This geographic performance gap reflects subtle differences in patient demographics, imaging equipment, skin care practices, and lesion presentation patterns across international clinical settings.
Artifact Impact on Accuracy Image artifacts including hair, pen markings, ulceration, and imaging source institution decreased accuracy in complex ways that varied by the underlying diagnosis. Some artifacts degraded performance uniformly across diagnoses while others showed differential effects, suggesting that robustness to artifacts must be addressed in a diagnosis-specific manner during algorithm development.
Metadata Benefits Algorithms in the task that incorporated clinical metadata (patient age, sex, anatomical location) showed modestly better performance than image-only algorithms on some categories. However, metadata alone was insufficient to close the performance gap created by distribution shifts, indicating that robust image-based learning remains essential.
The 'Not Trained' Category Problem AI systems have a fundamental inability to know what they don't know. When shown images of diseases absent from training data, algorithms cannot flag uncertainty - they must assign the image to one of their trained categories. The challenge specifically tested this failure mode, with results raising serious safety concerns.
47% False Malignant Rate For the top 25 submitted algorithms, 47.1% of images from the NT (not trained) category were misclassified as malignant diagnoses. This means that nearly half of patients presenting with unusual or rare conditions not represented in the training dataset would receive a malignant prediction, potentially leading to unnecessary biopsy or patient anxiety.
Human vs. Algorithm on NT Images Expert dermatologists correctly identified NT images as unusual or outside normal classification in 26% of cases, compared to only 6% for the top algorithms. This crucial gap shows that human experts maintain epistemic uncertainty and can recognize unfamiliar presentations, while AI systems confidently misclassify them.
Trained Category Comparison For disease categories included in training, algorithms generally performed better than or comparably to expert dermatologists in most diagnostic categories, except for actinic keratoses where performance was similar. This context-dependent superiority reinforces that AI excels within its training distribution but fails unpredictably outside it.
Safety Requirements for Deployment The findings identify specific safety requirements that must be met before AI skin cancer classifiers can be safely deployed clinically. Systems must demonstrate performance on data that matches the intended use population, must handle out-of-distribution images gracefully, and must be regularly validated as patient populations and imaging practices evolve.
Necessary Biopsy Rate Concern The high false malignant rate for NT images would translate into unnecessary biopsies at scale. Biopsies carry risks (infection, scarring, patient anxiety, healthcare costs) that must be weighed against diagnostic benefit. An AI system that triggers unnecessary biopsies for nearly half of unusual presentations would fail a clinical risk-benefit analysis.
Regulatory Testing Standards Current regulatory frameworks for AI medical devices do not adequately address the out-of-distribution failure mode identified in this study. The challenge results provide concrete evidence to support development of regulatory testing standards that require evaluation on distribution-shifted data and disease categories outside training.
Dermatologist Shortage Context The global shortage of expert dermatologists (ranging from 0.33 per 100,000 in the USA to 6.6 in Germany) creates genuine pressure to deploy AI tools. This finding reframes the goal from replacing dermatologists to developing AI that performs safely as a screening and triage tool, with human expert confirmation for all positive findings.
Open-Set Recognition The NT category failure requires development of open-set recognition systems that can identify when a test image is out of distribution and communicate uncertainty rather than forcing a confident misclassification. Anomaly detection, conformal prediction, and calibrated uncertainty quantification are active research areas that could address this safety gap.
Diverse Training Data Closing the geographic generalization gap requires deliberate inclusion of training images from diverse global populations, skin phototypes, imaging hardware, and clinical settings. The ISIC archive's continued expansion with international contributions and the establishment of minimum training diversity standards would improve clinical generalizability.
Prospective Validation Standards The challenge results provide a template for rigorous AI evaluation that should be adopted as a standard before clinical deployment. Pre-deployment evaluation should include geographically diverse test data, out-of-distribution cases, and comparison to dermatologists in matched head-to-head scenarios reflecting intended clinical use.
Iterative Post-Market Surveillance Even algorithms that pass rigorous pre-deployment evaluation require ongoing post-market surveillance to detect performance drift as clinical practice evolves, new disease categories emerge, or new imaging hardware is introduced. Regulatory frameworks should mandate performance monitoring through representative patient sample testing at defined intervals.