Scholarlink

Scientific integrity initiatives and funding requirements have motivated open access sharing of neuroimaging datasets that often include T1-weighted images. These images have voxels representing the face, which can sometimes be used to identify participants. Researchers often use skull-stripping tools to remove image voxels representing skull and facial features to comply with human subject privacy regulations. To ensure that no facial features are present in the skull-stripped images and that no voxels representing brain tissue are unintentionally removed during skull-stripping, time consuming and cumbersome visual inspection of the skull-stripped images is necessary to evaluate the risk for re-identification and data quality. Here, we describe an automated program that accurately identifies recognizable facial features in the skull-stripped images and detects loss of image voxels representing brain tissue to support inspection of data quality when sharing neuroimaging data. Specifically, a multi-kernel 3D Convolutional Neural Network (CNN) model with an inception module demonstrated a 95.49% accuracy in identifying recognizable facial features and a 97.63% accuracy in detecting the loss of brain tissue voxels. The training dataset and trained models are available online at https://dyslexia.computing.clemson.edu/QC_tool/ .

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

Deep Learning Methods to Evaluate Privacy and Quality of Skull-Stripped Brain Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Li Luo,
  • Rishikesh V. Phatangare,
  • James Z. Wang,
  • Kenneth I. Vaden,
  • Mark A. Eckert

摘要

Scientific integrity initiatives and funding requirements have motivated open access sharing of neuroimaging datasets that often include T1-weighted images. These images have voxels representing the face, which can sometimes be used to identify participants. Researchers often use skull-stripping tools to remove image voxels representing skull and facial features to comply with human subject privacy regulations. To ensure that no facial features are present in the skull-stripped images and that no voxels representing brain tissue are unintentionally removed during skull-stripping, time consuming and cumbersome visual inspection of the skull-stripped images is necessary to evaluate the risk for re-identification and data quality. Here, we describe an automated program that accurately identifies recognizable facial features in the skull-stripped images and detects loss of image voxels representing brain tissue to support inspection of data quality when sharing neuroimaging data. Specifically, a multi-kernel 3D Convolutional Neural Network (CNN) model with an inception module demonstrated a 95.49% accuracy in identifying recognizable facial features and a 97.63% accuracy in detecting the loss of brain tissue voxels. The training dataset and trained models are available online at https://dyslexia.computing.clemson.edu/QC_tool/ .