Research On Intelligent Prostate Cancer Detection Algorithm In Mp-MRI Images

Prostate cancer is a malignant tumor that occurs in prostate of men.Clinically,magnetic resonance(MR)imaging is mainly used to diagnose prostate cancer.The traditional manual dianosing method is time consuming and vulnerable to subjective factors.Computer-aided diagnosis systems can effectively improve the efficiency and accuracy of physician diagnosis.However,due to the small size and blurred contour of prostate cancer lesions,automatic detection of prostate cancer in MR images is quite challenging.The quantity of samples in the dataset is limited,for the reason that it’s quite difficult to build a prostate cancer dataset for deep learning.So a pipeline intelligent detection algorithm of prostate cancer based on self-supervised learning is proposed to make full use of limited data.The algorithm consists of three parts: registration of prostate MR images,prostate segmentation and prostate cancer lesion localization.First,registration between different series is done to find the spatial relation between the series.Second,the prostate segmentation module is uesd to segment the prostate so that we can limit the lesion inside the prostate and remove irrelevant background.Thirdly,the prostate cancer lesion localization module is uesd to segment the lesion and obtain the mask of the suspected nodule.Due to the limited samples in the dataset and abundant unlabeled data,a self-supervised method is adopted,which can make full use of unlabeled data and improve the performance of the algorithm.However,the pipline algorithm is vulnerable to accumulative error.So an end-to-end intelligent detection algorithm of prostate cancer is proposed.The system merge the task of prostate segmentation and prostate cancer lesion localization into one network,which can do multitask learning.After registration between different series,a co-trained network is used to segment prostate and prostate cancer lesion in parallel.By multitask learning,the network can learn better feature representation,which helps the network improve its performance.Finally,the pipeline algorithm achieves recall of 92.54% with FPPP of 0.6111 on the test set,while the end-to-end algorithm achieves recall of 89.18% with FPPP of 0.66427.It costs pipeline algorithm 4.21 s to predict a sample,while it costs the end-to-end algorithm 0.42 s.It can be seen that the pipeline algorithm is more accurate while the end-to-end algorithm is more efficient.Both can basically meet the clinical requirements.