Enhancing the Reliability of Medical Image Classification and Segmentation via Self-Supervised and Uncertainty-Aware Learning
- Publication Type:
- Thesis
- Issue Date:
- 2026
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Medical imaging plays a critical role in the diagnosis of diseases such as cerebral infarction, brain tumors, and lung tumors. However, two fundamental challenges limit the reliability of computational medical image analysis: severe class imbalance in classification tasks and pervasive label noise in medical image segmentation. In classification, pathological cases are typically underrepresented, leading to biased learning and reduced sensitivity to clinically significant abnormalities. This issue is further aggravated by substantial variability in pathological appearances. In segmentation, label noise arises from annotation variability, ambiguous anatomical boundaries, and human errors, resulting in inaccurate lesion localization and boundary delineation.
This thesis proposes two complementary methodologies to address these challenges. First, a self-supervised representation learning framework is developed to improve classification robustness under class imbalance, augmented by a Self-Calibrated Feature Denoising strategy that enhances minority-class feature stability without modifying the backbone network. Second, a noise-robust segmentation framework termed Label Uncertainty Transformation is introduced, integrating pixel-wise uncertainty estimation, entropy-based sample selection, and instance-dependent label transition modeling. Extensive experiments on private and public datasets demonstrate improved robustness, accuracy, and generalization across diverse clinical scenarios.
Please use this identifier to cite or link to this item: