Computer Aided Detection And Diagnosis Of Prostate Cancer Based On Deep Learning

Prostate cancer(PCa)is the most commonly diagnosed non-skin cancer in men and the second most common cancer in the world.Early detection,diagnosis and proper treatment are crucial for reducing mortality.Among imaging technologies,multi-parameter magnetic imaging(mp-MRI)is demonstrated to be the most effective method for PCa detection and diagnosis,which can significantly improve the detection rate and accuracy,and effectively guide the biopsy.Therefore,automated,fast and accurate computer-aided detection and diagnosis(CAD)of prostate cancer based on mp-MRI has become a research focus in the field of intelligent medicine for decades.Existing CADs typically use multiple independent modules for image registration,multimodal feature fusion,cancer probability map generation,and cancer localization and grading,each of which is optimized separately.In addition,these CADs differ in the handcrafted feature/representation of PCa.On the one hand,traditional handcrafted features and modules are less robust.On the other hand,separate optimization ignores error tolerance between modules,and the error will be accumulated among them.In recent years,deep learning has achieved great success in the medical field.Its characteristics including end-toend training,weakly-supervised learning,and automatic feature learning provide insights for the detection and diagnosis of PCa.This thesis investigates the application of deep learning in the detection and diagnosis of PCa,focusing on novel deep neural networks(DNNs)based CAD of PCa,and robust and efficient optimization methods.Specifically,the research contents in this thesis are as follows:(1)The optimization of existing CADs has to be guided by pixel-level labels.However,the PCa annotations are based on biopsys,and the collected data has only image-level labels to indicate whether there is cancer or not,but lacks pixel-level label for position and contour.An CAD of PCa based on weakly-supervised DNN was proposed.The system can learn to generate a reasonable cancer probability map for cancer localization by optimizing an imagelevel classification only.Meanwhile,two identical DNNs are utilized to process the mp-MRI images of two different modalities.To further improve the performance,a new loss function is designed to force the CAD to learn PCa-related features from data of different modalities while suppress those irrelated features.Experimental results show that the weaklysupervised CAD of PCa can learn relatively accurate cancer features and predict reasonable pixel-level cancer probability maps only through image-level labeling and multi-parameter feature fusion.(2)Aiming at the problem of error accumulation caused by separate optimization of modules,an end-to-end joint detection and diagnosis system for PCa is proposed.The system replaces the modules of mp-MRI image registration and prostate segmentation with a DNN and combines it with the weakly-supervised CAD of PCa to form a complete CAD of PCa.By applying error back propagation,all functional modules of the CAD are optimized in an end-to-end fashion,resulting in detected and well-registered prostate images of different modalities,and a reasonable cancer probability map for cancer localization and classification.Experimental results show that the end-to-end learning of CAD of PCa can improve mutual error tolerance among modules and significantly improve the overall performance of detection and diagnosis of PCa.(3)Medical data,especially positive data,is extremely rare and involves patient privacy,while the training of complex DNNs requires massive data.A mp-MRI image synthesis method based on generative adversarial network(GAN)is proposed.Unlike previous data augmentation methods,the method models the joint distribution of mp-MRI data in the lowdimensional manifold,and generates new unseen data by randomly sampling in the distribution.A semi-supervised training strategy is proposed to ensure the correct paired relationship between different modalities and the visual realism of the generated data.A stitchlayer is introduced to reduce the complexity of generation under unsupervised conditions.Auxiliary distance optimization is proposed to guarantee clinically meaningful cancerous visual patterns.Both the doctor's verification results and the objective experimental results show that the synthesis method can effectively generate mp-MRI data with varity,correct paired relationship and cilinically meaningful visual patterns...