Loading your page…
Dermatology Practical & Conceptual
By Ney Michel Lituma Villamar and Natasha Juliana Vaccaro Zambrano.
Retrospective artificial intelligence studies in dermoscopy require independent validation and transparent reporting before clinical translation. Computational feasibility is also relevant where high-end hardware is unavailable.
Ward brings these articles into the same app as your cards and questions. One free account for Ward, Suture, Osler and Prognia.
To evaluate the diagnostic accuracy and computational feasibility of a lightweight EfficientNet-V2-S melanoma classifier and to test the locked original model in a geographically independent dataset.
The development analysis used 9,645 BCN20000 images from 2,436 lesions, split by lesion into training, validation, and held-out test partitions. The threshold (0.91015625) was selected only on the validation partition as the maximum sensitivity with specificity >=0.90. The original checkpoint was then evaluated without test-time augmentation, retraining, fine-tuning, recalibration, or threshold re-estimation on a prespecified DERM12345 subset comprising 11,230 images from 1,425 patients. Confidence intervals used 2,000 cluster-bootstrap replicates by lesion for the internal test and by patient for external validation.
The internal test set contained 933 images from 244 lesions. AUC was 0.831 (95% CI, 0.771-0.883); sensitivity was 0.612 (0.517-0.700), and specificity was 0.881 (0.823-0.934). The external cohort contained 400 melanoma and 10,830 non-melanoma images. External AUC was 0.923 (0.896-0.946); sensitivity was 0.625 (0.540-0.697), specificity 0.978 (0.974-0.982), positive predictive value 0.513 (0.420-0.601), negative predictive value 0.986 (0.982-0.990), and balanced accuracy 0.802 (0.759-0.838). External calibration was poor (intercept -2.934; slope 0.601). The selected checkpoint was epoch 7; training required 47.6 minutes, with 936 MB peak allocated CUDA memory.
EfficientNet-V2-S training and inference were feasible on consumer-grade hardware, and discrimination remained high during geographic external validation. Nevertheless, the locked operating point had moderate sensitivity and substantial miscalibration. The model is not established as a rule-out, autonomous triage, or clinical decision system; prospective workflow, reader, calibration, and subgroup studies remain necessary.
Published in Dermatology Practical & Conceptual. The abstract is shown as published and remains the work of its authors; read the full article at the source.
Prognia lists this article for reference and education. It is not medical advice. Always read the original and follow current guidelines and local protocols before acting on it.
