MRI Segmentation Of Prostate Peripheral Zone Using Deep Convolutional Neural Networks

Doctors in community hospital have a high recognition with the rate of error diagnostic,due to lack of the clinical experience.In central hospital,Reading enormous amount of medical imaging make heavy work for physician.The algorithm of deep convolutional neural networks based on big data has a strong advantage in pattern recognition since 2012.And this method has already been successfully applied in computer vision,image processing and medical image analysis by scholars and researchers.AI interpretation can help doctor to relieve from heavy works of reading MP-MRI images,give support to doctor on prostate tumor diagnosis and improve specificity and sensitivity of clinical diagnosis.And,Medical image segmentation plays a key role in Computer-Aided Diagnosis system of prostate diseases,providing an essential technical for pathological analysis and stage of disease.According to research statistics that there are about 70% of Prostate Cancer(PCa)has been occurred in Peripheral Zone(PZ).Segmentation of prostate peripheral zone directly affects the quality of multi-modal image registration and tumor image recognition.The main works of this thesis are summarized as fellows:1.The thesis proposed a novel data augmentation method using Generative Adversarial Networks for prostate images.The modified Deep Convolutional Generative Adversarial Network(DCGAN)could generate the realistic-looking synthetic MRI prostate images,the total number of dataset expanded more than one fold.And,experiment show that the algorithm with data augmentation and transfer learning boosts the performance efficiently for prostate tumor classification.The proposed method improve the diversity of image and the size of MP-MRI prostate dataset.Adversarial learning has been shown to produce state of the art results for data fitting.2.The thesis presented an automatic segmentation of the prostate peripheral zone on T2 weighted(T2w)images is a necessary clinical applications for prostate cancer diagnosis.(1)Without adequate prostate peripheral zone images with ground truth,the proposed joint learning loss function based on adversarial learning could improve network’s performance of data fitting.The joint adversarial loss can solve T2 w smallsample problems.The proposed method improve the accuracy of object recognition and accelerate forward and backward propagation by transfer learning.(2)However,the low contrast,blur contour,and significantly varies of shape serious challenges to accurate segmentation of PZ.For extracting and encoding multi-level features about peripheral zone more effectively,multi-scale dilated convolution(MDC)and pooling block(MPB)were embedded in the baseline model U-Net.The modified network is a trade-off between high level semantic information extracting and recovery of low level feature maps,which is useful for segmentation.(3)Then,the thesis integrated a soft attention mechanism to focus on the salient features useful for segmentation of PZ,and suppress irrelevant information and tell the generator where to locate for PZ.In the end,experiments on 991 T2 w images of 142 patients demonstrates that the modified generative adversarial network based on multi-scale features enhancement is more robustness.The thesis about segmentation of prostate peripheral zone provide research value and clinical significant.