Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study evaluates four deep learning architectures—ViT-B/16, Swin-S, ConvNeXt-S, and EfficientNetV2-S—for the classification of dermoscopic images of skin neoplasms, comparing their performance across various classification schemes. The findings reveal that while the binary classification stage achieves high ROC-AUC scores internally, there is a significant drop in performance when applied to independent clinical datasets, highlighting a generalization gap. Notably, the two-stage cascade approach improves macro F1 scores by effectively recovering malignant lesions misclassified as benign, demonstrating the importance of sensitivity control in clinical applications.
The cascade classification approach significantly enhances sensitivity control, outperforming traditional single-stage methods in identifying malignant skin lesions.
Purpose. To compare deep learning architectures and classification schemes for dermoscopic images of skin neoplasms and assess their generalization on transfer from open international datasets to independent clinical datasets of Russian practice. Methods. Four architectures (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) were compared in three schemes: binary (malignant/benign), single-stage four-class (benign, MEL, SCC, BCC), and a two-stage cascade (binary triage, then three-class differentiation MEL/SCC/BCC). All models used ImageNet-pretrained weights and a single augmentation protocol on aggregated open ISIC Archive data, and were evaluated on an internal held-out sample and two clinical datasets (Melanoscope AI mobile system; Sechenov University). Results. Internally the binary stage attains ROC-AUC 0.952-0.966; on Sechenov University it drops to 0.797-0.893, sensitivity to 0.53-0.67, and ECE rises from 0.02 to 0.27-0.39 with underestimation of malignancy, quantifying a generalization gap in ranking and calibration. Paired tests confirm one inter-architecture result on clinical data: the deficit of ViT-B/16 at the binary stage (p<0.05); at the differentiation stage no architecture has a proven advantage. The cascade raises macro F1 over single-stage four-class classification for most architectures, but significantly only for ViT-B/16, by recovering malignant lesions assigned to the dominant benign class. On ISIC MILK10k, direct 11-class classification yields mean-class sensitivity 0.525. Conclusion. A tunable triage threshold gives sensitivity control not attainable in standard single-stage (argmax) classification and better reproduces clinical differential-diagnosis logic. The persistent generalization gap mandates external clinical validation and recalibration before deployment.