Ki-67 Counting Application Based On Optical Scanning Image

In the field of domestic pathology,currently the diagnosis is mainly based on pathologists with rich experience.However,the number of pathologists is small and the work efficiency is low.With the development of artificial intelligence machine learning and other fields,there is also a great possibility for pathological diagnosis to be intelligent.In order to improve the work efficiency of pathologists and the accuracy of Ki67 scores,this thesis improves the target detection algorithm based on Ki-67 images of breast cancer.The related work and improvements based on the Faster R-CNN network are as follows:(1)Since Ki-67 does not have a public data set,the large-resolution Ki-67 slide image provided by the hospital is segmented to generate a patch with a size of 1024×1024as the data set.At the same time,in order to reduce the workload of labeling,an iterative training method is used to increase training samples until the training requirements are met.(2)The backbone network is ResNeXt-101,then change the 7×7 convolution kernel to three 3×3 convolution kernels connected,which enhances the nonlinearity and complexity of the network while reducing the amount of parameters.At the same time,a 2×2 maximum pooling layer is added before the 1×1 convolution kernel whose step length of the residual module branch is 2,and the step length of the 1×1 convolution kernel is changed to 1 to reduce the loss of feature map information.(3)Aiming at the feature fusion method,a bottom-up and top-down two-way fusion method is used to fuse the feature maps,ensuring that the low-level feature maps contain high-level semantic information,and the high-level semantic information also contains information of low-level feature maps.After that,the Non-local model was added,allowing the network to learn the location information of the cells.The recall and accuracy have been improved in this way.(4)For the location loss function of the RCNN network,this thesis introduces a Balanced L1 loss function instead of Smooth L1 Loss.The Balanced L1 loss function improves the gradient contribution of small loss samples,and balances the learning between simple samples and difficult samples.At the same time,a post-processing method is added to the prediction end to solve the problem of two different types of detection results in the same cell.In the Ki-67 test images of breast cancer,the F1 score of positive cancer cells reached 95.2%,and the F1 score of negative cancer cells reached91.1%.Based on the actual application,this paper designs a set of end-to-end process for the whole silde images.In the real scene of breast cancer Ki-67 stained digital image slices,it provides an effective detection system.The system can improve the efficiency and accuracy of the doctor’s diagnosis.