{"title":"Brain Tumor Segmentation using MRI Images by Optimized U-Net","authors":"D. Ramya, C. Lakshmi","doi":"10.1109/ICNWC57852.2023.10127271","DOIUrl":null,"url":null,"abstract":"Segmenting tumor in the brain is a challenging process undertaken by the surgeon to assess and locate the tumor location in the MRI images. To overcome this constraint, an improved U-Net architecture for use in the BraTS20 and BraTS21 challenge’s brain tumor segmentation problem is proposed. The accuracy has been improved by modifying the loss function. Comprehensive ablation research to investigate Deep Supervision loss, Cross-Entropy, Decoder Attention, and Residual Connections to determine the best model architecture and learning schedule is performed. Multiple convolutional channels have been experimented with, and post-processing techniques to find the ideal spot for the U-Net encoder’s depth have also been undertaken. The proposed technique outperforms every U-Net variant and produces superior outcomes while incurring a minimal loss.","PeriodicalId":197525,"journal":{"name":"2023 International Conference on Networking and Communications (ICNWC)","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2023-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2023 International Conference on Networking and Communications (ICNWC)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ICNWC57852.2023.10127271","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
Segmenting tumor in the brain is a challenging process undertaken by the surgeon to assess and locate the tumor location in the MRI images. To overcome this constraint, an improved U-Net architecture for use in the BraTS20 and BraTS21 challenge’s brain tumor segmentation problem is proposed. The accuracy has been improved by modifying the loss function. Comprehensive ablation research to investigate Deep Supervision loss, Cross-Entropy, Decoder Attention, and Residual Connections to determine the best model architecture and learning schedule is performed. Multiple convolutional channels have been experimented with, and post-processing techniques to find the ideal spot for the U-Net encoder’s depth have also been undertaken. The proposed technique outperforms every U-Net variant and produces superior outcomes while incurring a minimal loss.