Research And Application Of Deep Learning For Medical Image Computer-aided Diagnosis

Medical image is an indispensable tool in clinical diagnosis.But to interpret medical images and present diagnoses are labor-intensive.An efficient computer-aided diagnosis system is urgently needed to help clinicians make better interpretations and decisions.Latest advances in deep learning have enabled us to rethink computer-aided diagnostic methods based on medical images.Although the clinical practice of computer-aided diagnosis based on deep learning is being accepted,the lack of large datasets with trusted labels is considered to be one of the biggest challenges in the successful deployment of deep learning methods in medical image computer-aided diagnosis applications.This dissertation takes breast ultrasound computer-aided diagnosis as the breakthrough point,and focus on tumor detection,few sample learning,noise label learning,and confidence calibration in breast ultrasound images.The main contributions and innovations are summarized as follows:Firstly,to solve the problem of detecting region of tumor(ROI)in breast ultrasound images,this dissertation analyzes several existing state-of-the-art object detection methods,and studies how to use deep learning to detect tumors in breast ultrasound images.Since there is no open source dataset for breast ultrasound images,in conjunction with the ultrasound specialists from Sichuan Provincial People's Hospital,this dissertation collects a new dataset of ultrasound breast images,and have them manually labeled by experienced clinicians.This dissertation employ the state-of-the-art deep learning based detection methods to locate tumor regions in breast ultrasound images and systematically evaluates them on newly collected dataset,and establish benchmarks for the newly collected dataset.Secondly,to solve the ROI classification problem of breast ultrasound images with noise labels,this dissertation proposes a new noise label classification method(COLCNet).The BI-RADS rating of the breast ultrasound image indicates the possibility of tumor malignancy.It is naturally suitable to be transformed into a label distribution method,and optimize the noise labeling problem.Based on this,COLC-Net proposes a soft label for breast ultrasound images.In addition,different classifiers can generate different decision boundaries with different learning capabilities.In order to update the soft label of breast ultrasound,COLC-Net uses network collaboration framework.In each batch of training,distill excellent knowledge from the two networks and dynamically update the soft labels to prevent overfitting noise labels and strengthen the learning of accurate labels.Thirdly,to address the noisy label problem in training complete breast tumor classification models,this dissertation proposes an effective approach called noise filter network(NF-Net).Specifically,NF-Net incorporating two softmax layers for classification,which significantly slows the training speed of deep models and hence prevents them from memorizing the training data of noisy labels.However,Double-softmax also slows the learning speed of accurate labels.In order to strengthen the effect of clean labels,we design a teacher-student framework for distilling the knowledge of clean labels.In this distillation model,in contrast to the student network using Double-softmax,the teacher network uses a single softmax layer to predict the benign and malignant labels of the training data,and uses the distillation loss function to perform knowledge distillation between the teacher and student network.Furthermore,a class balance regularization is incorporated over the predictions of the teacher network.Specifically,by maximizing the entropy of the marginal class distribution,the training samples with clean labels are preferred to be uniformly distributed across each class.The class balance regularization can improve the prediction quality of the teacher network and hence enhance that of the student network through the distillation loss.Fourthly,in order to solve the classification problem of benign / malignant tumors in breast ultrasound images with few samples,this dissertation proposes a siamese integrated learning framework(SIMK-Net)which integrates the domain knowledge of clean labels and noisy labels.In this framework,BI-RADS domain knowledge is proposed as a discrete distribution between different rating,and two class tasks are proposed.One task is based on the biopsy results.Another task is based on label correction of the BIRADS score,in which soft label is dynamically corrected by minimizing the mean square error loss between the soft label and the network prediction.With the accuracy improvement of soft labels,the network can learn richer features from the professional ultrasound doctor,improve the generalization ability,and thus improve the discrimination ability of the biopsy label classifier.The two classifiers help each other improve their capabilities.SIMK-Net learn more feature representations from different medical expertise,which can improve the overall ability of the model.Fifth,in order to solve the problem that the computer aided diagnosis system based on breast ultrasound cannot give reliable confidence,this dissertation proposes a voting-based confidence calibration method,which combines network training and post-processing.By using the correlation between the confidence and the cross entropy loss of the network,this method proposes a knowledge discrimination risk network and select the samples with high confidence in each batch to update parameter.Thus improving the recognition accuracy and relative reliability of the network.Then,bases on BI-RADS rating,calibrating the confidence boundaries with validation sets.Finally,proposing a confidence vote calibration method,dividing the training set into different subsets,voting the output results through the trained sub-model,and making final calibration for the confidence with boundary calibration.