Stroke is a serious disease that causes high mortality and disability rates worldwide. Early and accurate imaging diagnosis is used clinically to achieve early prevention and timely treatment. However, traditional imaging diagnostic methods have a strong dependence on the knowledge and experience of doctors, which can easily miss unclear lesion features. To this end, a novel image segmentation model CS-SegNet has been proposed, aimed at automatically segmenting lesion areas in stroke CT images to assist in diagnosis. CS-SegNet is based on the UNet-Resnet50 architecture and introduces Channel and Spatial Attention (CASA) modules in the down sampling stage to enhance its ability to extract contextual information from key regions; in the up sampling stage, combined with the RDSConv module, residual learning and dense connections are used to optimize convolution operations, improve feature expression ability, and enhance segmentation accuracy in complex backgrounds; the skip connection part introduces a multi-scale channel attention (MSCA) module, which integrates low-level detail information and high-level semantic information to improve the accuracy and consistency of segmentation results. The experimental results showed that CS-SegNet achieved segmentation accuracy, average intersection to union ratio, and recall rate of 99.79%, 91.52%, and 93.83%, respectively, which improved the performance of UNet Resnet50 basic network by 0.14%, 5.11%, and 4.05%, and performed the best in multiple comparative experiments. Compared with existing mainstream models, this model has effectiveness, good segmentation accuracy, and learning ability in stroke lesion segmentation.