Analyzing Overfitting Under Class Imbalance in Neural Networks for Image Segmentation

Class imbalance poses a challenge for developing unbiased, accurate predictive models. In particular, in image segmentation neural networks may overfit to the foreground samples from small structures, which are often heavily under-represented training set, leading poor generalization. this study, we provide new insights on problem of overfitting under class by inspecting network behavior. We find empirically that when with limited data and strong imbalance, at test time distribution logit activations shift across decision boundary, while well-represented seem unaffected. This bias leads systematic under-segmentation structures. phenomenon is consistently observed different databases, tasks architectures. To tackle problem, introduce asymmetric variants popular loss functions regularization techniques including large margin loss, focal adversarial training, mixup augmentation, explicitly designed counter classes. Extensive experiments conducted several challenging tasks. Our results demonstrate proposed modifications objective function can lead significantly improved accuracy compared baselines alternative approaches.