Narrowing the regional attention imbalance in medical image segmentation via feature decorrelation

Abstract

Convolutional neural networks with U-shaped architectures are widely used in medical image segmentation. However, their performance is often limited by imbalanced regional attention caused by interference from irrelevant features within localized receptive fields. To overcome this limitation, FDU-Net is proposed as a novel U-Net-based model that incorporates a feature decorrelation strategy. Specifically, FDU-Net introduces a feature decorrelation method that extracts multiple groups of features from the encoder and optimizes sample weights to reduce internal feature correlations, thereby minimizing the interference from irrelevant features. Comprehensive experiments on diverse medical imaging datasets show that FDU-Net achieves superior evaluation scores and finer segmentation results, outperforming state-of-the-art methods.