改进Swin Transformer骨干网络的病理图像分割方法

Pathological Image Segmentation Method Based on Improved Swin Transformer Backbone Network

在人工手动计数乳腺癌病理图像的有丝分裂结果过程中,由于形态相近的细胞存在干扰,目标细胞难以手动分割标记,从而限制乳腺癌分级诊疗效率和准确性的提高。为了解决以上问题,以近年来在计算机视觉多个子领域获得SOTA的Swin Transformer算法为基础,基于Swin Transformer进行改进,并与Unet Decoder进行组合,最后提出一种新的乳腺癌有丝分裂图像语义分割网络,有效的定位到有丝分裂细胞区域,并利用后处理进行个数统计。对现有语义分割特征提取算法Swin Transformer基本结构进行优化,为了适应有丝分裂细胞小目标区域,通过构建跨尺度稠密连接,将图像中的纹理特征加权融合到深层网络中,并设计了一种卷积核尺寸选择的注意力机制进行跨尺度稠密连接的构建,最后使用了一种加权的L1损失函数,保证目标图像准确性,进而实现对目标有效分割。在ICPR14数据集上进行评估,实验结果表明上述网络实际训练中表现出更为优越性能,在MIOU指标上达到了87.01,在Dice指标上达到了87.18,能准确高效分割乳腺癌有丝分裂病理图像。

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 vision 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 experimental 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 pathological images.