Research On The Model Of Glioma Automatic Segmentation Network Based On Deep Learning

Objective: The application of imageology in the diagnosis and treatment of glioma has a good prospect.Because manual tumor segmentation is time-consuming and has poor repeatability,and image segmentation plays an important role in the application of imageology,this study proposes an automatic segmentation network model of glioma based on deep learning(GCCUNet),and introduces Dice Similarity Coefficient(DSC),sensitivity and Intersection over Union,Io U)to evaluate the results.Methods: The preoperative head magnetic resonance imaging(Magnetic Resonance Imaging,MRI)images of 162 patients with gliomas confirmed by pathology after resection of intracerebral space occupying lesions in Shengjing Hospital affiliated to China Medical University from January 2015 to June 2022 were collected.Based on U-shaped network,using background removal and image amplification techniques,and introducing DR path,GCA module and MI path,GCAUnet,takes 21546 magnetic resonance images as training set to improve the generalization ability of GCAUNet model,and then selects 36 images as verification set to optimize parameters.Finally,182 magnetic resonance images are selected to test the model,and dice similarity coefficient,intersection ratio and sensitivity are introduced as evaluation indexes,manual segmentation is compared as the gold standard,and finally compared with several current segmentation networks.Results: The dice similarity coefficient(Similarity Coefficient,DSC)of the brain glioma automatic segmentation system constructed in this study is 85.82%,91.54% and 83.81%respectively,which is better than the current segmentation network models.Conclusion: In this study,an automatic segmentation system for glioma is proposed,which has higher segmentation accuracy than other segmentation network models,is closer to the gold standard of manual segmentation,and has a good application prospect.