MSFA-Net: Multi-scale feature aggregation and attention-enhanced U-Net for microscopic hyperspectral pathology images segmentation

Abstract

Pathological diagnosis is regarded as the gold standard for early gastric cancer detection. Automatic segmentation efficiently delineates lesion areas, which are essential for identifying and diagnosing these lesions. Current segmentation methods that rely on grayscale or RGB images are hindered by limited information, whereas advancements in microscopic hyperspectral imaging technology offer a novel perspective for early gastric cancer diagnosis. This technology provides rich spatial and spectral information that effectively reflects the chemical composition and physical state of tissues, thereby enhancing the identification of cancerous regions. U-Net has marked a significant advancement in segmentation networks and has demonstrated promising results in hyperspectral pathology images segmentation tasks. However, the substantial semantic gap between its encoder and decoder presents challenges in addressing complex lesion areas. To address this problem, a multi-scale feature fusion and attention-enhanced U-Net (MSFA-Net) is proposed which optimized for reduce semantic gaps, and applied to the segmentation of precancerous lesions in gastric cancer using microscopic hyperspectral pathology images. The Cross-Stage Feature Fusion (CSFF) module is designed to accurately merge encoder-level features, alongside Multi-Scale Integrated Convolution (MSIConv) extracts multi-scale features, effectively bridging the semantic gap between the encoder and decoder. Additionally, the Adaptive-weighted Attention (AWA) module is designed to optimize the fusion of encoder and decoder features, further enhancing the recovery of image details. Experimental results demonstrate that the model performs well on both Intestinal Metaplasia (IM) and Gastric Intraepithelial Neoplasia (GIN) stages of precancerous gastric cancer.