Semi-Supervised Medical Image Segmentation via Frequency Attention with DCT and Data Exchange: The FAS-Net Approach

Abstract

In the domain of medical imaging, precise segmentation of organs or pathological regions is vital for disease diagnosis and various clinical practices. The inherent complexity of medical images poses challenges in procuring pixel-level exact annotations. Consequently, researchers have explored semi-supervised approaches to medical image segmentation, needing minimal pixel labeling for effective segmentation. Despite noteworthy advancements in this field, there remains potential for boosting model performance. This study proposes a new FAS-Net architecture for enhanced semi supervised medical image segmentation. This model is based on the Mean Theacher structure, and a frequency attention guidance module is designed based on the discrete cosine transform. An input data augmentation strategy for data exchange is introduced to improve the efficiency of semi supervised medical image segmentation. The frequency attention module empowers the segmentation network to mitigate non-relevant image regions while accentuating critical features. The data exchange strategy counteracts data distribution inconsistencies and empirical discrepancies by randomly selecting and interchanging segments between pairs of images. The efficacy of the FAS-Net model, when combined with the U-Net segmentation architecture, is demonstrated through validation on two datasets: ACDC for cardiac condition assessment and BraTS2019 for brain tumor assessment. Experimental outcomes demonstrate that FAS-Net surpasses existing state-of-the-art semi-supervised medical image segmentation techniques across these datasets.