Research On Classification Of Breast Cancer Recurrence Risk Grade Based On Digital Pathological Images

With the development of medicine,the prevention and treatment of breast cancer has attracted more and more attention.Breast cancer has become a common disease worldwide,and it is the deadliest and most common cancer among women.Early detection of cancer is very important,because with the early diagnosis of the disease,the chances of cure and recovery increase.For patients with early breast cancer,although they are in the same clinical stage or pathological grade,their postoperative recurrence risk levels are still different.For different early patients with different recurrence risk levels,the treatment methods used are also different.Early patients with a high risk of postoperative recurrence can use adjuvant chemotherapy to prevent tumor recurrence and metastasis.Early breast cancer patients with low risk of recurrence can receive endocrine therapy.Therefore,accurately assessing the recurrence risk level of patients with early breast cancer and formulating an accurate treatment plan for them is the current focus of research by clinicians.With the rapid development of deep learning in recent years,convolutional neural networks have been introduced in many fields.At the same time,histopathological examination is regarded as the "gold standard" of diagnosis,and quantitative analysis can be performed quickly through the computer-aided diagnosis system.In summary,this article uses deep learning to apply digital pathology images to classify breast cancer patients’ recurrence risk levels.The main tasks include:(1)Establishment of breast cancer histomathological image annotation data set: the most core part in the application of deep learning technology is data set.Because only on a certain number of data sets can the neural network model complete the training and fully learn the characteristics of the image.This paper mainly collects pathology images from public data sets(TCGA),and clinicians from Beijing Metadata Genomics Co.,Ltd make corresponding labels,and collect clinical data sets and match corresponding labels to meet the actual needs of the research.(2)Normalization of the color of breast cancer tissue images: Due to the high heterogeneity of breast cancer,the morphology of its histopathological images is different.This paper adopts a supervised classification framework to calculate the method of image-specific staining matrix,using the lαβ color space.The characteristics of channels that are not directly related to each other.Color correction is achieved by selecting a suitable target image and applying its characteristics to another image.(3)Research on breast cancer pathological image block prediction recurrence classification algorithm: For the uniquely labeled breast cancer pathological image in this article,combined with clinical practice,a fully connected layer is added to the training network to merge the learned features and reduce the divergence of features.The impact of target classification;secondly,add the Centerloss function on the basis of the softmax Loss function to reduce the distance between classes.This enables the model to classify the recurrence risk level of early breast cancer patients with a higher accuracy rate.(4)Research on the introduction of migration learning methods to predict recurrence level classification algorithms: In view of the actual situation of the small data set in this article,a pre-training set is proposed,and different methods are used to fine-tune the four networks of VGG19,Res Net50,Inception V3 and Dense Net169 through migration learning.,It solves the problem of pathological image shortage,and also improves the classification accuracy of the convolutional neural network.