Research And Implementation Of Computer Aided Prostate Cancer Detection Algorithm For Multimodality MRI

Prostate cancer has a high incidence in men.Magnetic resonance imaging is currently the most commonly used imaging diagnostic method.However,manual interpretation of magnetic resonance data is time-consuming and labor-intensive.Automatic,rapid and accurate computer-aided detection and diagnosis algorithms for prostate cancer are an important research direction in the new generation of artificial intelligence medical field due to the need to shorten the reading time and reduce the pressure on the reading physicians.The research of this paper is to study and implement deep learning algorithms for automatic segmentation and classification of prostate cancer lesions in magnetic resonance images,so as to assist doctors in finding the location of cancer lesions and grading of cancer lesions.The paper mainly does the following work:1.Design and implementation of prostate cancer segmentation algorithm based on modality fusion and shape learning.In this paper,the characteristics of prostate magnetic resonance imaging are considered in the design of deep learning network,and how to reasonably integrate multimodality information and utilize edge shape characteristics is considered.A two-stage cascaded network(MFSL-NET)based on modality fusion and shape learning was designed for prostate cancer segmentation.The network adopts a cascade structure,which cascades the two networks proposed in this paper:1)Modality fusion network,which is a dual-flow CNN,and makes two modality flows interact in two dimensions.The spatial attention module and channel attention module are introduced to extract supervisory information so that information can be shared and interactive between modalities.2)Shape learning network is a multi-task learning network that combines semantic segmentation and edge detection to recognize shape and edge information,so that shape information can be better learned and recognized among all kinds of fused information.In this paper,the design choice of the algorithm is proved by ablation experiments and comparative experiments,and its performance is compared with that of existing methods on an open dataset.The results show that the dice similarity coefficient is improved by 3.6%,which confirms the effectiveness of the proposed algorithm.2.Design and implementation of prostate cancer classification algorithm based on soft thresholding.In this paper,a soft thresholding based classification algorithm is proposed.A deep learning network based on ResNet was established to classify clinically significant prostate cancer and non-clinically significant prostate cancer using multimodality magnetic resonance images.In view of the multi-noise characteristics of MRI images,soft thresholding used in signal denoising algorithm is introduced for 2D medical image classification.Soft thresholding and deep learning are combined to eliminate the noise-related information,that is,the redundant information irrelevant to the classification task is suppressed,and a more accurate classification result is obtained,the accuracy rate reached 86.2%.