In whole slice images, renal tissues exhibit irregular shapes, significant size variations, and similar textures and structures across different tissue types, accompanied by the challenge of class imbalance. To address these issues, a lightweight kidney pathological tissue segmentation network, ASRMU-Net, is proposed. First, a spatial reconstruction unit (SSRU) is introduced in the shallow layers of the network to capture spatial information from different renal tissues using average and maximum values. Through gated mechanisms and convolutional operations, spatial features are adaptively reconstructed, and useful features are enhanced via cross-recombination. In the middle layers, an ASRM module is designed to enhance feature representation by reconstructing and fusing spatial and channel features. In the deep layers, a channel reconstruction unit (CSRU) is integrated. This unit combines adaptive channel splitting, compression, and depthwise separable convolution to effectively fuse high-dimensional and low-dimensional features, adaptively reconstruct them, and distinguish tissues with similar textures and structures. Finally, an improved loss function is applied to optimize the model and reduce the impact of class imbalance. The enhanced network achieves MDice and MIoU scores of 85.4% and 74.8% on the interstitial fibrosis dataset, and 96.1% and 92.4% on the AIDPATH dataset. The results demonstrate that the improved network achieves higher segmentation accuracy with fewer parameters compared to other medical segmentation models.