Deep Learning Classifier And Its Interpretability Research Work Basedon Breast Ultrasound Image

Breast cancer is a high mortality rate cancer which seriously endangers women's health.Many factors may lead to it.Ultrasound detection is simple,painless,noninvasive and non-radiative.It is the main method in breast tumors diagnosis stage.The ouput of ultrasound detection contains breast ultrasound image and report.Meanwhile,the breast imaging report and data system(BI-RADS)developed by the American Society of Radiology have strict specifications for the ultrasound report.Using ultrasound data and doctor's professional knowledge to diagnose breast cancer highly depends on the doctor's individual ability which is too much subjective.The computer aided system(CAD)combined with artificial intelligence and medical diagnosis technology,will help detecting and classifying of lesions and reduce the rate of missed diagnosis and improve doctors' work.effectiveness.The earlier version of computer aided diagnosis system first extracts the texture and shape features of the image by artificial or algorithm methods,then forms the feature vector,finally trains the machine learning model with the feature vector.In recent years,the deep learning theory represented by “deep neural network”(DNN)has gradually matured.DNN can represent high spatial dimension.The features extracted by DNN have stronger generalization performance than the artificial method does,it is continuously applying in health and medical fields now.In order to explore the application of deep learning in the breast cancer diagnosis,this paper designed a CNN-based classifier on the breast ultrasound dataset.According to the standard lesion segmentation of professional physicians,the accuracy of the training set ACC is 0.7866,the validation set is 0.8.It is worth noting that there is great imbalance between positive and negative samples in ultrasound samples,which leads to a low rate of sensitivity.My paper has tried two methods to optimize the loss function and achieved good results.The ACC of the Bias-Loss method on validation set up to 0.8222,sensitivity is 0.8667,specificity is 0.7556,Am-Loss method ACC is 0.86,sensitivity is 0.9565,specificity is 0.78,both effectively improve accuracy and sensitivity.Applying deep learning methods for decision-making,especially in the medical images diagnosis,the interpretability of decision-making and output process is not enough.The differences between deep CAD and evidence-based diagnosis limits the trust of doctors and patients.In order to improve the interpretability of the deep learning algorithm and promote it's application,My paper attempts to interpret the model through heat map and semantic regression.The heat map effectively found the region algorithm evaluated,the semantic regression discovered the important features of the model.Finally,my paper integrates the above work and develops a client program for breast ultrasound aided diagnosis system with automatic segmentation,classification,and interpretation functions.