Anchored Supervised Contrastive Learning for Long-Tailed Medical Image Regression

Imbalanced data distributions are common issues in image regression tasks, particularly in medical diagnosis where the inherent nature of the data is imbalanced. In these situations, the minority data is often of significant importance. Recent studies have shown that supervised contrastive learning can significantly enhance performance in long-tailed recognition tasks by balancing the quantity of positive samples. However, these approaches cannot be directly applied to regression tasks that need to predict continuous values. Meanwhile, they also cannot effectively address the issue of inadequate feature space uniformity, evident in that samples from the many-shot region occupy much more feature space than those from the few-shot region, despite having label space of the same length. To address these issues, we propose an Anchored supervised contrastive learning method for long-tailed Medical Image Regression (Anchored-MIR). Specifically, we design a Regression Supervised Contrastive (RSCon) loss, which consists of two components: the Adaptive Positive Sample Selection (APSS) strategy to tackle the scarcity of samples in the few-shot region, and adaptive label distance to incorporate label distance relationships for better adaptation to the image regression task. Then, we devise an Anchored Supervised Contrastive (ASCon) loss to forcibly expand the feature distribution and improve the separability of samples in the few-shot region. Finally, to further exploit the limited information from samples in the few-shot region, we also design a soft-reg loss, which is re-weighted by a Triangular Kernel Weighting (TKW) function that mines the similarity in the relationships between adjacent data samples. Experiments on the bone-age assessment task demonstrate that our Anchored-MIR outperforms existing state-ofthe-art long-tailed regression methods.