Prostate Segmentation And Cancer Detection Based On Multi-parameter MRI

Prostate cancer is the second most common malignant tumor for men in the world,and the incidence rate keeps increasing every year,which is called "the second largest killer of men".Early diagnosis of prostate cancer and reasonable treatment can effectively alleviate the suffering of patients and improve the chance of survival.Magnetic Resonance Imaging(MRI),as a non-invasive and low-cost diagnostic technique,has become an important method for radiologists to diagnose prostate cancer and is widely used in clinical practice.Computer-Aided Diagnosis(CAD)can assist doctors to find lesions through imaging and improve the accuracy of diagnosis.Gland segmentation and cancer detection for prostate MRI have become important research content.This thesis analyzes the way radiologists deal with prostate MRI images,proposes prostate segmentation and cancer detection algorithms based on deep neural networks,and obtains practical demonstrations on a large number of data samples through experiments.The main research contents include:(1)Propose a prostate segmentation algorithm based on polymorphic geometric constraints.The network framework consists of two stages: 3D coarse segmentation and 2D fine segmentation: the first stage takes the 3D segmentation network V-Net as the backbone,adds prediction branches of key points and boundaries,and uses polymorphic geometric information to constrain the gland shape as a whole;based on rough segmentation,the attention mechanism is used to strengthen the correlation information between different slices,and the segmentation of small gland slices at both ends is optimized.At the same time,this thesis also introduces the shape compactness loss and the two-dimensional weighted Dice loss to ensure the overall shape of the gland and guide the network training to pay more attention to the target area and improve the segmentation performance.Experiments show that the predicted values of Dice coefficient,accuracy,Hausdorff distance,and average surface distance of the gland segmentation algorithm are 92.65%,93.56%,0.154 mm,and 0,.847 mm,respectively,which are overall better than the comparison algorithms in this field.(2)Propose a multi-parameter MRI prostate cancer detection algorithm.In this thesis,by modeling the signs of cancer lesions that radiologists pay attention to in the clinical diagnosis of prostate multi-parameter MRI images,we analyze the effects of different parameters of MRI on prostate cancer lesions and design an appropriate feature fusion encoder to balance spatial information and semantic information.At the same time,to make the training and learning of the network more robust,this thesis adopts a task-driven loss function,which assigns different weights to the suspicious areas of cancer foci.The experimental results show that the algorithm’s Dice coefficient,accuracy,sensitivity,and specificity are 68.59%,71.25%,71.76%,and 99.89%,respectively,which achieves better performance than related algorithms.(3)Propose a multicenter MRI prostate cancer detection algorithm.Aiming at the problem of domain generalization in prostate multi-parameter MRI data,the distribution of different data domains in the feature space is converged through feature center domain alignment training,thereby improving the network generalization ability and improving the performance of the network on unknown data domains.promote.The final experiments show that compared with the network before the feature center domain alignment training,the algorithm has greatly improved the Dice coefficient of the unknown data domain on the task of cancer detection.The above studies show that the algorithm proposed in this thesis can achieve the task of segmentation of glandular regions in prostate MRI images and detection of cancer foci,which not only improves the accuracy of doctors in diagnosing prostate cancer but also provides patients with glands and cancer in the follow-up radiotherapy process.The relative regional position of the focus can achieve efficient diagnosis and precise radiotherapy.