LRG-UNet:A Lightweight Randomized Grouped Feature Aggregation Enhanced UNet for Rectal Cancer Gland Segmentation

Colorectal adenocarcinoma is a common malignancy of the digestive system, posing significant threats to multiple aspects of patients’ health, quality of life, and survival. Hematoxylin and eosin (H&E) stained slides are widely utilized in pathology for tissue staining, and semantic segmentation enables the precise delineation of various tissue components, such as cancer cells, normal glands, and stroma. This process assists pathologists in accurately identifying the location, extent, and type of tissues involved in the tumor, thereby enhancing the accuracy and efficiency of diagnosis. To accurately segment glandular regions in H&E-stained colorectal adenocarcinoma tissue images while reducing computational costs, this paper proposes a U-Net model based on multi-axis focused attention and randomly grouped feature aggregation, named LRG-UNet. The model introduces three key improvements: First, in the skip connections, a randomly grouped feature aggregation block is used to fuse low-level features at each stage, which are then input to the corresponding decoder layers. Second, multi-axis focused attention is employed for global feature extraction, enhancing the model’s capacity for global feature modeling. Finally, deep supervision is utilized as the training strategy. Experimental results on the public GlaS dataset for colorectal adenocarcinoma segmentation demonstrate that the algorithm achieves Intersection over Union (IoU), Dice similarity coefficient, Precision, and Recall of $\mathbf{7 1. 2 3 \%, ~} 81.84 \%, \mathbf{9 0. 8 2 \%}$, and $\mathbf{7 8. 0 9 \%}$, respectively. Compared to other segmentation methods, this algorithm achieves superior performance with fewer parameters.