Medical Image Segmentation Based On Deep Learning Method

Medical image is a reflection of the internal structure of the human body,and it is one of the main bases of modern medical diagnosis.Medical image processing tasks mainly include image detection,image segmentation,image registration,and image fusion.At present,the object of medical image segmentation is mainly the image of various human organs,tissues,and nuclei.The task of medical image segmentation is to divide the image into several regions according to the similarities or differences between regions.In recent years,many researchers have continuously explored and proposed a variety of medical image segmentation methods and technologies,and have proposed many methods,including threshold-based segmentation methods,region-based segmentation methods,and edge detection-based segmentation methods.The model based on traditional machine learning methods such as decision tree,random forest,and clustering algorithm achieves better results in image segmentation.Because the traditional machine learning methods mainly rely on feature engineering,the expression ability of extracted features is limited,so the performance is limited.In recent years,deep learning methods,especially the correlation methods based on convolutional neural networks,have good feature recognition ability and generally better performance than traditional machine learning methods in medical image segmentation and other fields.Medical image segmentation methods based on deep learning have attracted more and more attention and applications.Aiming at the characteristics and application requirements of medical images and pathological data,especially the retinal fundus blood vessel image,human epithelial cell image,and nuclear pathological image of various human organs,this dissertation studies the segmentation method of medical images based on deep learning.In this paper,to realize the high quality of the medical image segmentation as the goal,from the medical image data preprocessing,image visual feature recognition and extraction,feature weighting method of adaptive selection,post-processing and optimization,etc.,of the segmentation results of conducting thorough research,propose some innovative theories and methods.This dissertation analyzes and studies the existing digital imaging technology,summarizes the research status,the commonly used methods,and performance evaluation indexes of medical image segmentation based on deep learning,and point out that the performance indexes of existing segmentation methods such as accuracy and accuracy are not high enough,and the segmentation results are prone to be interfered by lesions and equipment noise.It is difficult to distinguish the organ images with complex morphology and structure.To solve the problems in human retinal fundus image segmentation,such as the small number of blood vessel pixels,the imbalance of positive and negative sample categories,the easy fracture of fine blood vessels in the segmentation,and the interference of lesions and equipment noise,the research on retinal fundus vascular image segmentation method based on the improved residual network and attention mechanism was carried out.A novel retinal vascular image segmentation model AR-SA U-Net was proposed.This model is based on the classic medical image segmentation model U-Net and can improve the segmentation quality of microvessels,suppress the interference factors caused by lesions or devices,and effectively segment retinal fundus vascular images by combining the residual network module with cavity convolution,sc SE-based attention mechanism,and Inception network module.The proposed model was trained,validated,and tested on three open datasets: DRIVE,STARE,and CHASE＿DB1,and multiple ablation experiments were performed.At the same time,the generalization ability of the model was tested and verified on a liver dataset and optic disc dataset of glaucoma.Experimental results show that the performance of the proposed model is better than that of many other models in recent years.Deep learning-based nuclear segmentation methods for pathological images have problems of over-segmentation and under-segmentation,and the segmentation ability is insufficient in the case of nuclear adhesion and overlapping,and the segmented nuclear contour is easy to break.To solve these problems,this dissertation firstly proposed an attention mechanism based on the spatial and channel mapping table method and combined with recurrent convolution residual module,proposed a nuclear contour segmentation model RCSAU-Net for nuclear contour segmentation in pathological images.Then,based on the proposed RCSAU-Net model and adversarial generation network,nuclear entity segmentation was carried out.In addition,the multi-task learning method and the proposed post-processing method based on the voting mechanism to dynamically adjust the boundary weights were used to separate and merge the nucleus contour and entity respectively,which solved the problem of adhesion and overlapping nuclei difficult to separate.Experimental results on Mo Nu Seg and Pan Nuke datasets show that the segmentation performance of the proposed multi-task deep adversity-learning method exceeds that of BRP-NET,Hover-Net and Star Dist.Due to the complex structure of medical images,the difficulty of data annotation,and the lack of labeled data,the existing convolutional neural network-based methods still have shortcomings.For example,the segmentation result precision is not high enough,the segmentation is not fine enough,and the existing model is mainly proposed for specific datasets,so the generalization ability is weak.Based on Transformer’s powerful global context information representation ability and CNN’s local feature extraction ability,this dissertation carries out research on semantic segmentation methods of medical images based on multi-scale Transformer and CNN.The fusion method of Transformer and CNN,the feature fusion method of multi-scale Trans-CNN module,and the confidence optimization method of segmentation result cascade of side output results are proposed.Experiments on dataset DRIVE,ISBI2018 Cell,and Mo Nu Seg show that the proposed method has good performance and generalization ability.