Attention-Guided and Noise-Resistant Learning for Robust Medical Image Segmentation

Accurate segmentation of anatomical structures or pathological lesions from medical images is crucial for reliable disease diagnosis and organ morphometry assessments in clinical practice. However, the presence of artifacts, occlusions, and uneven brightness introduced during image acquisition can present significant complexity in achieving segmentation. In this article, we propose a robust attention-guided and noise-resistant (AGNR) framework that jointly integrates attention-guided mechanisms and noise-resistance capabilities within a transformer-based U-shaped encoder-decoder to address the challenge posed by noise interference in the context of medical image segmentation. In particular, the AGNR framework derives benefits from the coordinated interaction between its spatial attention and channel attention components. Spatial attention enables our model to prioritize salient regions within the medical images, which effectively mitigates the adverse effects of occlusions and artifacts. Concurrently, channel attention dynamically accentuates relevant appearance characteristics, thus addressing challenges stemming from uneven brightness and variations in organ texture. Within each encoding stage, a noise-resistant semantic distillation (NRSD) module is thoughtfully inserted to enhance the robustness of the model against noise interference by modeling semantic uncertainties inherent based on a teacher-student manner. Furthermore, our approach additionally incorporates advanced information bottleneck (IB) techniques into the decoding stage, which aims to reduce the imprecision and ambiguity within feature representation. Extensive experiments on multiple benchmark datasets with diverse imaging modalities consistently demonstrate the superiority of AGNR over state-of-the-art baselines.