کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Embedding-Consistent Contrastive Learning: A Robust Approach for Imbalanced Classification

نویسندگان :

Sobhan Siamak¹ Eghbal Mansoori²

1- دانشگاه شیراز 2- دانشگاه شیراز

کلمات کلیدی :

contrastive learning،deep learning،imbalanced classification،representation learning،medical image analysis

چکیده :

This research introduces Embedding-Consistent Contrastive Learning (EC-CL), a novel framework that leverages contrastive learning to alleviate the adverse effects of class imbalance in medical image analysis. EC-CL architecturally enforces a structured latent geometry, moving beyond simple instance discrimination to ensure both intra-class compactness and inter-class separation within the embedding manifold through a specialized cosine-based objective function. The framework's performance was analytically evaluated on a comprehensive suite of six medical imaging sub-datasets from the MedMNIST benchmark—including BloodMNIST, DermaMNIST, OrganCMNIST, OrganSMNIST, RetinaMNIST, and PneumoniaMNIST. Experimental results demonstrate that EC-CL consistently outperformed state-of-the-art counterparts, achieving an average improvement of 4.3% in AUC and 3.8% in accuracy on multi-class tasks. A notable advantage is that it secured these superior results using lower-resolution inputs, underscoring its computational efficiency. This work establishes EC-CL as an exceptional and principled paradigm for imbalanced classification, effectively bridging robust representation learning with the demands of real-world clinical deep learning applications.