Research On Multi-task Deep Learning For Abdominal MR Image Segmentation

With the rapid development of artificial intelligence technology,deep learning have gradually become the mainstream research in medical image analysis with their powerful abstraction capabilities.Through the analysis of medical images,it is possible to accurately locate the lesion,formulate a suitable radiotherapy plan,and achieve personalized and precise treatment.But deep learning has high requirements on the quality and quantity of training data,as well as computing resources.The data volume of medical images is usually small and the imaging quality is poor.These factors bring challenges to medical image analysis using deep learning methods.In MRI,low-field-strength MRI has a larger scanning space,smaller equipment,and can move the treatment bed according to clinical needs.It is suitable for assisting doctors to perform chemoradiotherapy.How to accurately segment the region of interest in low-field-strength magnetic resonance images is a question worthy of study.Cervical cancer shows high signal and clear tissue outline on T2-weighted MRI.The accurate delineation of cervical cancer areas in magnetic resonance imaging is the premise and main basis for doctors to make radiotherapy plans.This article aims at the problem of severe noise and low data volume in low-field-strength magnetic resonance gastric segmentation,and the need to segment the cervical cancer region in magnetic resonance in the clinic.The segmentation algorithm is designed to assist doctors Clinical diagnosis and treatment.We mainly completed the following three tasks:1.In view of the problem that the low-field MRI image is severely noisy and has a partial volume effect,which affects the accurate segmentation of the stomach region,this chapter proposes a low-field MRI stomach segmentation model based on multi-task deep learning.Firstly,for the problem of serious noise in low field strength MR images,BM3 D algorithm is used to denoise and enhance it.Then build a unified Convolutional Neural Network based on multi-task learning,use the denoised images as high-quality reference image,and use the manually delineated stomach label as supervision information to train the unified model.Since the network structure and network parameters of the image feature extraction part of the image enhancement branch and the image segmentation branch are shared,the features extracted by the two tasks can be complementary,and at the same time,the risk of network overfitting can be reduced and the effect of image segmentation can be improved.2.For the low-field-strength MR data,over-fitting problems that would occur when using 3D network model segmentation directly,this chapter proposes a low-field-strength 3D MR image stomach segmentation model based on transfer learning image enhancement.First,the trained Cycle GAN network is used to realize the conversion between high and low field strength MR images,and the high field strength MR images are converted into pseudo low field strength MR images with similar visual performance and image distribution as the low field strength MR images.Then use the generated low-field-strength MR image and the original low-field-strength MR image together as a training data to train the 3D Res-Unet segmentation network.The experimental results show that the cyclic generation of the adversarial network is used to learn the mapping relationship between the high and low field strength MR images.The method of training the segmentation network by those data can improve the generalization ability of the segmentation network and achieve a more ideal segmentation effect.3.In order to solve the problem that the thickness of the MR image scan layer of cervical cancer is large and the resolution between the layers is low,the 3D convolutional neural network is directly used for segmentation and the effect is poor.This chapter proposes a 3D magnetic resonance cervical cancer segmentation network model based on multi-view feature fusion.The core of the model is the proposed multi-view feature fusion module,which draws on the design ideas of the residual module and the Inception module,extracts the features of the input image blocks from different perspectives,and then uses the channel attention module to weight the features Fusion.The experimental results show that through the refined design feature extraction block,it can effectively solve the problem of low resolution between the layers caused by the large thickness of the scan layer of cervical cancer magnetic resonance images,and improve the segmentation effect of cervical cancer.