Pathology Image Analysis Software Based On IOS Mobile Client

In recent years,more and more applications of computer vision technology in real life have promoted the development of many fields such as security,transportation,and military.In the medical field,with the increasing digitization of various medical images,more and more researchers are focusing on the integration of computer vision and medical image.In the clinic,pathological diagnosis is called a significant standard for benign and malignant tumors.However,the complex and large-capacity features of pathological images lead to a heavy workload for professional pathologists.Therefore,computeraided diagnosis for pathological analysis is of great significance.Based on computer vision technology,this paper studies the segmentation and classification of breast cancer pathology images,and developed a pathological image analysis annotation software based on iOS mobile terminal.The main work includes two parts:(1)Aiming at the characteristics of pathological datasets of lymph node metastasis in breast cancer,we propose a structure-aware framework to analyze breast cancer metastases based on convolutional neural network.The research is divided into two parts: image segmentation for extracting lymphoid structure and image classification for determining lymph node metastasis stage.In image segmentation task,we first study and practice the common method of thresholding segmentation for extracting the region of interest in pathological images,and then we improve the experimental results by semantic segmentation based on fully convolution network,and obtaining more precise structures of the lymph region.In the image classification task,we designed a two-stage classification experiment because a pathology image has huge number of pixels.Firstly,we introduce lymph structure information to guide patch extraction for binary classification.We compare a transfer learning method and a fine tuning method without any pre-trained weight,and the performance shows the fine tuning model is superior which accuracy is 98.78% in test datasets.Secondly,we extract the features based on the classification results combined with the lymph structure,to stage the lesions with random forest method which ROC score is 0.8538.(2)We developed a pathology analysis annotation software based on the system of iOS.The trained classification model is integrated into the app,and makes predictions based on new input patch data.We analyzed the function and use of representative pathological annotation software.Take advantage of the good interaction of the mobile device,use the perception of finger movement instead of using mouse to mark,record the feature information with screen recording and audio recording,instead of keyboard typing.Optimizing the process of PC-side software requires using external devices frequently.