Research On Deep Learning Method For Abdominal Organs And Lesions Recognition

With the rapid development of artificial intelligence technology,in-depth learning has performed well in computer vision,natural language processing and other fields.A variety of applications of "artificial intelligence +" have begun to fall into the ground.Among them,the research of artificial intelligence in the medical field has important practical significance.Through in-deep learning,complex medical images are processed to assist clinicians to complete high-repetition work,which not only reduces the workload of doctors,but also brings better experience for patients.Starting from the actual clinical medical problems,in view of the practical problems in the process of abdominal prostate radiotherapy and colorectal endoscopy,this paper proposes the corresponding in-depth learning method of abdominal organ and lesion identification,which can assist doctors to complete the clinical diagnosis task efficiently and accurately and adjust the treatment plan in time.(1)A prostatic image registration method based on automatic segmentation and pelvic alignment is proposed to solve the problem of difficult selection of reference points and large registration deformation in the process of multi-modal prostatic image registration.Firstly,the region of interest in prostatic MRI and CT images is segmented.Then,the relative position information of prostatic organs and surrounding pelvis in the region of interest are used as a prior constraint,and the rigid anatomical structure of pelvis is used as a reference standard for rigid rough registration based on maximum mutual information.Then,according to the characteristics of deformation of prostatic organs in multimodal images,Demons method is used to fine-tune the results of rough registration.This method can obtain more accurate registration results under the condition of small deformation.(2)A SPVGG-based image classification method for colorectal endoscopy is proposed to solve the confusion between polyposis and other diseases in the process of inspecting colorectal endoscopy.In this method,according to the idea of self-step learning,the samples are divided into difficult and easy samples,by calculating the loss value of each sample's forward propagation in VGG network and comparing with the age parameters,and then the samples are gradually added into the training model from easy to difficult.SPVGG method can significantly improve the recognition rate of polyps on the basis of guaranteeing the classification accuracy of other categories.Finally,in order to assist enteroscope doctors to accurately determine whether polyps need surgical resection,we transfer the model trained in the colorectoscope image classification task as the encoder of the segmentation network,and successfully completes the polyp segmentation task.(3)A semi-supervised pre-3D prostatic image segmentation method based on generative adversarial network is proposed to solve the practical problem of prostatic MRI image segmentation in medical image processing,which is difficult for doctors to label manually and has few medical image labels.Firstly,the discriminator for generative adversarial network is replaced by the 3D U-net network.At the same time,a small amount of labeled data,unlabeled data and generated data are input into the discriminator respectively,and the weighted sum of the labeled data features and unlabeled data features extracted from the discriminator and the difference items of generated data features constitutes a new loss of the generator.This method can accurately locate the prostate under the condition of fewer labeled samples,and can segment the prostate organs accurately.