LEARNING CLASS-CONDITIONAL TEMPERATURE WITH ENTROPY ALIGNMENT FOR MEDICAL IMAGE CLASSIFICATION
Lists
Only you see this
Abstract
Medical image classification plays a pivotal role in computeraided diagnosis but remains challenged by the domain gap between natural and medical images and by severe class imbalance. Conventional cross-entropy losses often fail to address these issues simultaneously, leading to biased predictions and poor calibration. We propose Class-Conditional Temperature Entropy Align Loss (CCT-EAL), which integrates class-conditional temperature scaling with entropy alignment. The former adaptively rescales logits to mitigate imbalance, while the latter enforces distribution-level consistency to improve calibration and robustness under distribution shifts. Extensive experiments on five public medical benchmarks demonstrate improvements on most backbones and datasets. Compared with specialized medical models, CCT-EAL is efficient and generalizable, often matching or surpassing them under a modest training budget (10 epochs, single NVIDIA RTX 4090D). These results support CCT-EAL as a simple, effective component for more calibrated and class-balanced medical image classification. Our code will be available on GitHub.
BibTeX
@misc{pan2026learning,
author = {Weichao Pan and Xu Wang and Ruida Liu and Chengze Lv and Puyu Han and Luqi Gong},
title = {{LEARNING CLASS-CONDITIONAL TEMPERATURE WITH ENTROPY ALIGNMENT FOR MEDICAL IMAGE CLASSIFICATION}},
year = {2026},
doi = {10.36227/techrxiv.177015972.25096353/v1},
}Loading related content...