Research On Medical Image Segmentation Alogrithm Based On Machine Learning

With the improvement of computer hardware processing ability,machine learning algorithm has been applied to more and more fields.It has been mature in the field of natural image segmentation,but there are still some difficulties to be solved in the field of medical image segmentation.This dissertation studies the medical image segmentation methods based on machine learning.Its task is to segment lesions or other organs and tissues from medical images to provide assistance for doctors’ diagnosis.Medical image segmentation is a key step in clinical diagnosis.Quickly and accurately obtaining tissues that need further observation from scanned images can speed up the diagnosis and treatment.Due to the complexity of human organs and tissues,noise in medical equipment imaging and limited amount of medical data,it is difficult to achieve ideal results by directly using traditional natural image segmentation methods.Therefore,it is necessary to design some more accurate segmentation algorithms combined with some characteristics of medical images.In this dissertation,traditional image segmentation methods and image segmentation methods based on deep learning are studied.Their basic principles and progress in the field of medical images are introduced.The characteristics and segmentation difficulties of medical images and the basic structure of convolutional neural network used in this dissertation are deeply discussed.Also several classical convolutional neural networks and image segmentation networks are introduced.This dissertation mainly uses convolutional neural network to segment medical images.Firstly,it improves the general coding and decoding network,and then combines the advantages of other networks to improve the U-Net network.The algorithm proposed in this dissertation has achieved good results on three different medical image data sets.The main work of this dissertation is as follows:(1)This dissertation introduces and compares the traditional image segmentation methods and the image segmentation methods based on deep learning.It also explains why the traditional image segmentation algorithm is difficult to apply to medical image segmentation.Then it introduces three medical imaging methods:CT,MRI and ultrasound,and studies the difficulties of medical image segmentation and how to choose the appropriate preprocessing method.Finally,the basic structure of convolutional neural network,several classical convolutional neural networks and classical image segmentation networks are introduced.(2)For CT medical images,an improved coding and decoding network structure is proposed to segment the lung region from the chest CT image.The performance is improved to a certain extent by deepening the network.At the same time,batch normalization is introduced to accelerate the network convergence and obtain more accurate segmentation results.Experiments on the open chest CT image data set provided by Kaggle show that the deeper model segmentation effect is finer,and the introduced batch normalization can not only achieve the expected effect,but also accelerate the convergence of the network.(3)For MRI medical images,an improved U-Net segmentation algorithm is proposed.Taking the U-Net network as the framework,the residual module is combined with shallow features to make the MRI segmentation of brain tumors more precise.In order to further improve the efficiency of feature extraction and improve the performance of the model,attention mechanisms in space and channel are added in this dissertation.This attention module can make the network focus on more useful features at the spatial and channel levels at the same time,and weaken the weight of the features with small effects.At the same time,because the tumor accounts for less proportion of the whole image,this dissertation uses the weighted loss function to make the background account for less weight.In addition,preprocessing methods for medical images such as data enhancement are also adopted.Experiments on BraTS 2018 dataset show that the addition of residual network and attention mechanism improves the segmentation performance.(4)For ultrasound medical images,the U-Net network is applied to the segmentation of breast cancer ultrasound images,combined with batch normalization to improve the convergence effect.The network is optimized by dropout and finally verified by experiments.This structure has also achieved good results in ultrasound imaging of breast cancer.