SLRBC: self-supervised learning for red blood cell segmentation and classification with a biased contrastive loss
Yicheng Leng, Eung-Joo Lee · 2025
Within Red Blood Cell (RBC) image datasets, variations in cell morphology, such as schistocytes and teardrop-shaped cells, provide significant clinical information. Analyzing these shapes can provide valuable insights for diagnosis. Although recent advances in deep learning have notably improved image analysis for classification and segmentation tasks, particularly with large, labeled datasets, their effectiveness on RBC datasets can be limited by factors such as restricted data diversity and cases in which visually similar images are assigned different labels. To address these challenges, we propose a novel self-supervised contrastive learning framework that integrates negative pair selection with a biased contrastive loss. In specific, we incorporate an image similarity measure for negative pair selection in contrastive learning and employ a modified contrastive loss optimized for our framework. Using a publicly available RBC dataset, we validate our method, which enhances discrimination among different RBC types and improves classification performance. Experimental validation with various configurations demonstrates that our framework not only addresses challenges in RBC image analysis but also provides promising directions for applying self-supervised learning techniques in this domain.