Research On Multi-instance Pathological Image Classification Method Based On Deep Learning

As the intelligent development of soft hardware equipment,pathological imaging technology is also becoming increasingly mature,the analysis of pathological images has gradually turned into a research hot of international and internal,and is widely used in the prognostic analysis of lung cancer,breast cancer,colon cancer,esophageal cancer,celiac disease and other diseases.Accurate prognosis judgment is the hinge to the further therapy of sufferers.Therefore,accurate prediction of tumor development before surgery or neoadjuvant chemoradiotherapy is very important to formulate individualized treatment plans,including avoiding overtreatment and timely selection of alternative treatments.In addition,it can help patients avoid the risks and uncertain consequences of surgery.At present,the mainstream pathological image analysis models often use deep learning-based multi-instance learning methods to analyze pathological images,that is,the complete pathological image is first sampled in blocks,then each slice is processed separately,and finally,the patch-level analysis results are aggregated to obtain patientlevel predictions.However,due to the specificity of pathological images,these methods still have various problems.Therefore,this topic studies pathological images of various cancers,aiming to solve problems in specific fields and betterment the generalization and categorization capability of the model,so as to escort the development of precision smart medicine.The main work and contributions are as follows:(1)A multilayer perception(MLP)-based method for gated attention weight normalization(GAWN)is proposed.By decoupling the weight vector in the gated attention network and re-parameterizing it into a length scalar and a direction vector,thereby accelerating the convergence of stochastic gradient descent during model training.(2)A bilinear attention multi-scale feature fusion method(Bi FF)is designed to alleviate the problem of global information loss.By learning the bilinear attention distribution of pathological images,the mechanism makes full use of the designated sense of sight information and better integrates the global features provided by pathological slices with larger receptive fields and the detailed features provided by pathological slices with smaller receptive fields.(3)In addition,since the cancerous area in the pathological image is much less than the healthy area,the problem of imbalance of positive and negative instances is often caused after block sampling,therefore,we design a weighted loss function to alleviate this problem by simultaneously optimizing the instance-level and package-level loss functions.All in all,this thesis proposes a more advanced pathological image classification method based on deep learning and applies it to tumor prognosis research.Experimental results on a private locally advanced rectal cancer dataset and two publicly available breast cancer metastasis datasets show that the method can powerfully enhance the classification and generalization performance of the model.