Research On Medical Image Segmentation Method Based On Deep Learning

Since the late 20 th century,cancer has gradually become a major factor leading to increased mortality.Radiotherapy is one of the important means of cancer treatment.In the process of formulating radiotherapy plan,the key to avoid adverse effects after treatment is to accurately delineate the healthy organs at risk of radiation.Traditional manual sketching is done by experienced doctors.However,due to fatigue,manual sketching is time-consuming and error-prone.In this context,in order to better meet the needs of clinical application,this paper proposes an optimized U-Net image segmentation network and an image segmentation network based on fusion detection and segmentation.The main work is as follows:1.An optimized U-Net image segmentation model was designed.In the original U-Net network structure,Inception structure,residual structure and extrude excitation structure are added to retain the main structure of U-Net,and each sub-module is used to obtain more image feature information.It solves the problem that the accuracy of the segmentation results obtained by the original U-Net network is low.2.The image segmentation model with fusion detection and segmentation structure is designed.In this paper,a two-step network is used to detect the target of the image,and then the target image is segmented for the detected area.The 3D convolution operation is adopted to correspond the CT image data to the human anatomical features,so as to further enhance the ability of acquiring image feature information.It solves the problem of low recognition and poor segmentation effect of organs in small areas.No matter using the data set constructed in this paper or the public data set for testing,the overall image segmentation accuracy has been significantly improved.3.Constructed the PKU3 dataset.According to the requirements of clinical application,the organ segmentation data set of CT images in the head and neck region was constructed,which included 19 organs and tissues.It is stored in the form of CT value and spatial information.Solve the problem of data scarcity and annotation scarcity in the field of medical image.4.Results evaluation and comparison.On the PKU3 data set,the average DICE coefficient and the average 95% Hausdorff distance obtained by using the optimized U-Net segmentation network model are 78.226% and 8.789 mm,respectively.The average DICE coefficient and the average 95% Hausdorff distance obtained by using the fusion detection-segmentation network model are 80.382% and 6.876 mm,respectively.Compared with other current algorithms,the results show that the segmentation algorithm in this paper has higher image segmentation accuracy.