Work

Trustworthy DINOv2-based skin lesion classification with uncertainty quantification and robust learning from limited medical images

Year: 2026
Type: article
Abstract: Developing reliable automated medical image classification systems remains challenging because annotated medical datasets are often limited, lesion categories are highly imbalanced, and high-confidence errors can occur even when conventional classification performance is strong. This study presents a reliability-oriented skin lesion classification framework based on pretrained DINOv2 representations, Multi-Head Self-Attention (MHSA) feature refinement, a lightweight multilayer perceptron classifier, and uncertainty-aware inference. DINOv2 is used as a pretrained visual backbone, while the downstream classification components are trained in a supervised manner on a lesion-disjoint partition of the HAM10000 dataset. Monte Carlo Dropout is employed during inference to quantify predictive uncertainty and support selective prediction, while post-hoc temperature scaling is used for probability calibration. On the fixed 992-image test set, the representative DINOv2 + MHSA configuration achieved an accuracy of 82.96%, weighted F1-score of 0.8342, macro F1-score of 0.7502, macro AUC of 0.9601, and weighted AUC of 0.9551. Comparisons with ImageNet-pretrained ViT-B/16, Swin-Tiny, and EfficientNet-B0 provided additional context for conventional classification performance. The reliability analysis showed that uncertainty-based error detection achieved an AUROC of 0.8293 and risk–coverage analysis obtained an AURC of 0.0502. Temperature scaling reduced the Expected Calibration Error from 0.0911 to 0.0369, while the Brier Score decreased from 0.2651 to 0.2467. The MHSA ablation showed improved feature-space organization and ranking-based discrimination, although threshold-dependent classification metrics did not improve uniformly. Limited-data experiments further examined performance when only 10%, 25%, 50%, and 100% of the training data were available. Overall, the study evaluates DINOv2-based skin lesion classification from both predictive and reliability perspectives and emphasizes uncertainty estimation, calibration, selective prediction, representation analysis, and data efficiency rather than claiming uniformly superior classification accuracy.
Language: English
Subfield: Oncology
Field: Medicine
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7225912756
FWCI:
Cites: 25
Cited by:
Related to: 10