Research On Medical Image Segmentation Based On Deep Learning

Medical image segmentation is a key issue in determining whether medical images can provide a reliable basis for clinical diagnosis and treatment.It is also important in the analysis of biomedical images.Recently,because of the application of deep learning algorithms in medical image segmentation,medical image segmentation technology has made significant progress.At present,deep learning algorithms have been successfully applied to many medical image segmentation problems.However,for some rare diseases with a variety of image representations that lack large amounts of data support,data-dependent deep learning algorithms are difficult to perform.Aiming at this problem,this paper proposes a medical image segmentation framework based on deep learning.This framework uses the mathematical morphology tool constrains the segmentation model,and combines multiple segmentation models designed for different tasks in the same network for collaborative training.The morphological constraints of the target and the synergy between multiple models effectively solve the problem of lesion segmentation with insufficient data volume and complex image representation.Aiming at the lesions with complex image representation,this paper proposes an image segmentation objective function under the constraints of morphological features.Mathematical morphology is used to extract the geometric features such as the area and perimeter of the segmentation target from the mask and the output of the segmentation network.At the same time,a branch for predicting these geometric features is designed for the network;The features extracted from the label and the mask are used to build the morphological constraints in the objective function and improve the overall grasping ability of the network for the segmentation target.We validated this model on a neuroblastoma segmentation dataset consisting of CT scans of248 patients.The experimental results show that morphological constraints have significant gains on the neuroblastoma segmentation network.Aiming at the problem of medical image segmentation without a large amount of annotation data,this paper proposes a multi-task learning framework for collaboratively segmenting different targets.By relying on the differences and complementarities of different segmentation targets and model structures,the high-level information of the underlying information and the semantic information are effectively shared.For the cooperative segmentation of different structures in medical images,we use the method of sharing the underlying parameters to obtain the underlying features that are favorable for segmentation and to fuse the unshared high-level semantic information to utilize the complementarity between different segmentation targets.Experiments on multiple tasks of multiple organ segmentation and neuroblastoma segmentation in the abdomen have shown that different segmentation targets in medical images can be effectively complemented in our multitasking learning framework.