Research On Medical Image Segmentation Based On Lightweight Convolutional Neural Network

As an important auxiliary tool to clinical medical diagnosis,medical images play a vital role in the diagnosis of patients’ conditions and surgical treatment planning.However,medical images have a higher complexity compared to natural images.Medical image segmentation separates the target organ tissue from the complex medical image,and then enables medical experts to easily analyze the target,effectively reducing the complexity and difficulty of target analysis.How to accurately and efficiently achieve medical image segmentation has been an important topic in medical images.Based on the convolutional neural networks,this thesis presents a lightweight convolutional neural network-based medical image segmentation method,which can effectively reduce the number of network parameters and improve the speed of model operations.The proposed method uses a symmetric encoder-decoder architecture,where the encoding network is used to output a convolutional feature map of the image,and the decoding network is used to perform multiple upsampling of the feature map and output a semantic segmentation map of the image.This thesis focuses on the following aspects:(1)A study on liver segmentation method based on lightweight convolutional neural network.This study combines the 3D-UNet network with Ghost Module to lighten the3D-UNet and effectively reduce the number of network parameters while ensuring the image segmentation accuracy.The research method is validated by segmentation experiments on abdominal liver images,which show that the method can segment medical images more accurately,and the number of network parameters of the proposed method is smaller and more efficient.(2)A study on colon cancer segmentation method based on two-stage convolutional neural network.The method mentioned in(1)is not effective enough in segmentation of low-contrast,small lesion areas(such as colon cancer),where segmentation is more difficult.This research method combines target detection and semantic segmentation techniques in the field of computer vision,and designs a two-stage lightweight network for detection and segmentation to improve the segmentation accuracy of small lesion regions and reduce the number of parameters and computational effort.(3)Based on the integration of the above two tasks,this thesis designs and develops a medical image segmentation system.The system has the main functions of 3D medical image display,automatic image segmentation,segmentation result visualization,and so on,which meets the basic needs of doctors to view and analyze medical images,and can segment and highlight the target area of interest,thereby reducing the complexity of medical images and enabling the doctor concentrates on the target area and assists the doctor in clinical diagnosis.