Lightweight multi-scale attention group fusion structure for nuclei segmentation

The intricacy of segmentation is intensified by the morphological variability of cell nuclei. While, the U-Net model can achieve commendable outcomes in such contexts, it encounters difficulties, including semantic inconsistencies between the encoder and decoder, as well as an excessive number of parameters. To tackle these challenges, this study presents a lightweight multi-scale attention fusion (MAF) module intended to supplant conventional skip connections, with the goal of alleviating semantic discrepancies arising from multi-level downsampling and upsampling processes. Specifically, we establish multi-scale skip connections utilizing various attention mechanisms to enhance the flow of information. Furthermore, we introduce a deformable attention Condense (DACond) module designed to replace convolution operations, thereby reducing the overall parameter count and enhancing ability to learn the shape of the nucleus. The proposed model is designated as Lightweight Multi-scale Attention Fusion UNet (LMAF-UNet). Our LMAF-UNet exhibits enhanced segmentation performance across four publicly available datasets, while concurrently minimizing parameter size and computational complexity.