TrustworthinessSafetyTransparencycs.CV
Improving Predictive Confidence in Medical Imaging via Online Label Smoothing
Kushan Choudhury, Shubhrodeep Roy, Ankur Chanda, Shubhajit Biswas, Somenath Kuiry
October 22, 2025
BIO Web of Conferences
3 citations
Key Takeaways
-
Overconfidence is a real problem in medical AI: Deep learning models used for medical image classification often make overconfident predictions, which poses a safety risk in healthcare settings where clinicians need to know when a model is uncertain about its diagnosis.
-
Online Label Smoothing improves both accuracy and reliability: The study found that Online Label Smoothing (OLS) improved classification accuracy across multiple neural network architectures (ResNet-50, MobileNetV2, and VGG-19) while also making models better calibrated—meaning their confidence levels more accurately reflect their actual accuracy.
-
The method considers relationships between medical conditions: Unlike traditional label smoothing approaches that treat all wrong answers equally, OLS dynamically adjusts during training based on the model's own learning patterns, which better reflects how medical conditions can be related to one another.
-
Practical solution tested on real-world scale: The research was validated on RadImageNet, a large-scale medical imaging dataset, using commonly deployed architectures, suggesting this approach could realistically be implemented in clinical AI systems without requiring major infrastructure changes.
-
Better calibration supports trustworthy AI in healthcare: By producing more reliable confidence estimates alongside improved accuracy, OLS helps build AI systems that clinicians can better trust and interpret, which is essential for responsible deployment in medical settings where errors can have serious consequences.
Abstract
Deep learning models, especially convolutional neural networks, have achieved impressive results in medical image classification. However, these models often produce overconfident predictions, which can undermine their reliability in critical healthcare settings. While traditional label smoothing offers a simple way to reduce such overconfidence, it fails to consider relationships between classes by treating all non-target classes equally. In this study, we explore the use of Online Label Smoothing (OLS), a dynamic approach that adjusts soft labels throughout training based on the model's own prediction patterns. We evaluate OLS on the large-scale RadImageNet dataset using three widely used architectures: ResNet-50, MobileNetV2, and VGG-19. Our results show that OLS consistently improves both Top-1 and Top-5 classification accuracy compared to standard training methods, including hard labels, conventional label smoothing, and teacher-free knowledge distillation. In addition to accuracy gains, OLS leads to more compact and well-separated feature embeddings, indicating improved representation learning. These findings suggest that OLS not only strengthens predictive performance but also enhances calibration, making it a practical and effective solution for developing trustworthy AI systems in the medical imaging domain.
Accepted and presented in International Conference on Advancing Science and Technologies in Health Science
Paper Info
arXiv ID2510.20011
PublishedOctober 22, 2025
Citations
3
Relevance75%
VenueBIO Web of Conferences
TypeJournalArticle, Conference
Categorycs.CV
Responsible AI Topics
TrustworthinessSafetyTransparency
Links
Citation data from Semantic Scholar