Scholarlink

Accurate automatic segmentation of tumor in brain MRI scans is necessary for reproducible radiomic data extraction. Deep CNN models perform this task with reasonable accuracy. This paper proposes a new image preprocessing technique which further improves this accuracy. The pipeline consists of a Gaussian transformation of intensity (GTI) on T2-weighted images to selectively enhance the tumor bearing regions. This is coupled with an adaptive thresholding technique to localize the maximum peritumoral region. Performance was evaluated on both 2D and 3D UNet-based deep segmentation models. The Multimodal Brain Tumor Segmentation Challenge 2020 (BraTS-2020) dataset is taken for model training, validation, and testing. To assess the effectiveness of our pre-processing pipeline for both high-grade glioblastoma (HGG) and low-grade glioblastoma (LGG), BraTS 2018 and BraTS 2019 datasets have been categorized in HGG and LGG. Then, segmented results are analyzed separately for both HGG and LGG. Our segmentation results are compared with other reported models, such as 2D deep residual UNET, 3D-UNET, and nnU-Net. It is observed that the proposed data conditioning improves the performance of delineating tumors. This has been ratified by carrying out a paired t-test, a statistical significance test.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Data Conditioning to Improve Quality of Segmentation of Brain MRI Using Deep CNN Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Surajit Kundu,
  • Jayanta Mukhopadhyay,
  • Nishant Chakravorty,
  • Rimpa Basu Achari,
  • Santam Chakraborty

摘要

Accurate automatic segmentation of tumor in brain MRI scans is necessary for reproducible radiomic data extraction. Deep CNN models perform this task with reasonable accuracy. This paper proposes a new image preprocessing technique which further improves this accuracy. The pipeline consists of a Gaussian transformation of intensity (GTI) on T2-weighted images to selectively enhance the tumor bearing regions. This is coupled with an adaptive thresholding technique to localize the maximum peritumoral region. Performance was evaluated on both 2D and 3D UNet-based deep segmentation models. The Multimodal Brain Tumor Segmentation Challenge 2020 (BraTS-2020) dataset is taken for model training, validation, and testing. To assess the effectiveness of our pre-processing pipeline for both high-grade glioblastoma (HGG) and low-grade glioblastoma (LGG), BraTS 2018 and BraTS 2019 datasets have been categorized in HGG and LGG. Then, segmented results are analyzed separately for both HGG and LGG. Our segmentation results are compared with other reported models, such as 2D deep residual UNET, 3D-UNET, and nnU-Net. It is observed that the proposed data conditioning improves the performance of delineating tumors. This has been ratified by carrying out a paired t-test, a statistical significance test.