Computer-Aided Detection For Prostate Cancer Using Multi-Parametric Magnetic Resonance Imaging

Prostate cancer(PCa)is the second most diagnosed cancer in men all over the world.PCa growth is characterized by two main types of evolution:(i)the slow-growing tumours progress slowly and usually remain confined to the prostate gland;(ii)the fast-growing tumours metastasize from prostate gland to other organs,which might lead to incurable diseases.Therefore,early diagnosis and risk assessment play major roles in patient treatment and follow-up.In the last decades,new imaging techniques based on Magnetic Resonance Imaging(MRI)have been developed improving diagnosis.In practise,diagnosis can be a detected by multiple factors such as observer variability and visibility and complexity of the lesions.In this regard,computer-aided detection and computer-aided diagnosis systems are being designed to help radiologists in their clinical practice.Computer aided detection and diagnosis(CAD)system is designed to assist doctors to diagnose effectively and extra burden of doctors.At present,the vast majority of prostate cancer detection and positioning systems was divided into four dependent branches:1)Registration;2)Prostate segmentation;3)Feature extraction;4)Classification for true lesions.Obviously,step dependence will result in the loss of information,making the positioning and classification results inaccurate.And it uses mostly traditional image processing algorithms.This article closely follows the popular deep learning algorithm,applies the algorithm to medical images,and proposes an end-to-end neural network structure based on weakly supervision to solve lack of data labels with medical image and loss of information.The main work is as follows:i)Propose a weakly supervised end-to-end automated prostate cancer location system,and extend to cancer classification.The system can input complete ADC and T2w image pairs,and then segmentation in the network,as well as the location of cancer lesions and classification of cancer grades.The inconsistency loss were proposed to jointly optimize the response map generated by ADC and T2w images consistent.This loss function makes our proposed network structure have better classification and positioning result of prostate cancer diagnosis,and has achieved very good performance.ii)Aneural network based multi-modal TDN is proposed for image registration,apply TPS to get deformation image.The network structure can be applied before any neural network architecture to registration image.The reverse conduction uses the loss function of the later classification and positioning accuracy to joint optimize the registration net,not just based on mutual information,which can make the registration more accurate.