Pathological Image Segmentation Method Based on Improved Swin Transformer Backbone Network
In the process of manually counting the mitotic results of breast cancer pathological images,due to the interference of cells with similar morphology,it is difficult for target cells to manually segment markers,which limits the improvement of the efficiency and accuracy of grading diagnosis and treatment of breast cancer.In order to solve this problem,based on the Swin Transformer algorithm which has obtained SOTA in several subfields of computer vi-sion in recent years,this paper improves it based on Swin Transformer,and combines it with Unet Decoder.Finally,a new semantic segmentation network for breast cancer mitotic images is proposed,which effectively locates the mitotic cell region,and uses post-processing to count the number.In this paper,the basic structure of the existing semantic segmentation feature extraction algorithm Swin Transformer is optimized,ensuring the accuracy of the target image and achieving effective segmentation of the target.The evaluation was conducted on the ICPR14 dataset,and the experi-mental results show that the above network exhibited superior performance in actual training,reaching 87.01 on the MIOU index and 87.18 on the Dice index,which can accurately and efficiently segment breast cancer mitotic patho-logical images.
Deep learningSemantic segmentationAttention mechanismPathological image of breast cancerMi-tosis
NETL
NSTL
万方数据
