Mehran Zoravar

Topic

Entropy-Aware Skin Lesion Classification via Conformal Ensemble of Vision Transformers

Department of Mechanical Engineering

Date & location

• Thursday, August 28, 2025

• 10:00 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Homayoun Najjaran, Department of Mechanical Engineering, University of Victoria (Supervisor)

• Dr. Flavio Firmani, Department of Mechanical Engineering, UVic (Member) 

External Examiner

• Dr. Brandon Haworth, Department of Computer Science, University of Victoria 

Chair of Oral Examination

• Dr. Xiaodai Dong, Department of Electrical and Computer Engineering, UVic

Abstract

Uncertainty quantification is an inherent part of decision-making in various do mains, lending credibility and interpretability to predictive models. In medical image analysis, where decisions involve high stakes, the precise prediction becomes an imperative task. Skin lesion classification is a significant application in this domain, necessitating robust uncertainty handling to ensure diagnostic accuracy. This paper presents the Entropy-Aware Conformal Ensemble of Vision Transformers (EACE ViTs), which applied Vision Transformer (ViT) models, Generative Adversarial Net works (GANs) for efficient data augmentation, and entropy-based ensemble weightings in a synergistic manner to meet these requirements. The proposed framework addresses key challenges in the domain, including class imbalance and the prevalent problem of low-confidence predictions due to the intricate nature of skin lesions. A distinctive feature of this work is the adaptive entropy adjustment applied to the thresholds in Regularized Adaptive Prediction Sets (RAPS) and Adaptive Prediction Sets (APS) methods, which calibrates uncertainty based on model confidence. This kind of adjustment provides more adaptive and flexible prediction sets. When tested with the HAM10000 dataset, EACE-ViTs not only exhibits better performance in accuracy, coverage, and uncertainty metrics but also outperforms the baseline CNN based models as well as individual ViT models by a considerable margin. The results validate the potential of EACE-ViTs as a promising tool for accurate and reliable classification of skin lesions in important medical applications.