Hybrid U-net based on channel reconstruction and self-attention calibration bridge for weakly-supervised cell segmentation

Semantic segmentation of cellular images obtained via optical microscopy is a critical area in medical image analysis. However, the irregularity of cellular images on glass slides, along with excessive impurities, presents significant challenges for existing segmentation algorithms, leading to frequent semantic misclassification. This paper introduces a Hybrid U-Net architecture based on encoding channel reconstruction and self-attention calibration bridging to achieve weakly supervised extraction of cellular category information. The architecture, built upon a hybrid CNN and transformer module framework, first proposes channel reconstruction during the convolutional encoding phase, creating a latent space for channel information by expanding the dimensions. Subsequently, texture and global information within the latent space are compressed and reconstructed, effectively eliminating inter-channel redundancy at minimal additional cost. Additionally, a self-attention calibration bridging module is employed in the skip connections, where attention is extracted from shallow features in two dimensions. The proposed method is tested and validated on the publicly available SIPaKMeD dataset, outperforming other semantic segmentation networks in terms of mIoU and Dice scores.