Scholarlink

Portable ultra-low-field (uLF, i.e., 0.064T) magnetic resonance imaging (MRI) offers a solution to scarce radiological alternatives of resource-limited regions; however, in such regions, MRI system operators and radiologists are novices to the underrepresented modality. Therefore, automatic methods that confirm image acquisition of appropriate quality for diagnosis and segment and measure critical anatomical structures are required to support rural sites. This paper describes our approach to two tasks presented in the LISA 2024 Challenge: (1) quality assurance of pediatric low-field MRI data using DenseNet, and (2) segmentation of the bilateral hippocampi using nnUNet. As uLF MRI natively introduces unique image quality challenges, we trained seven DenseNet264 models, each designed to detect a specific artifact type: noise, zipper, positioning, banding, motion, contrast, and distortion. Similarly, as uLF MRI struggles to offer strong anatomical delineation, we enhanced challenge provided 0.064T low-field MRI scans using a custom Super-Field Network (SFNet) to generate high-quality Super-Field (SF) images, then trained an nnUNet model on these SF images with the LISA Challenge provided hippocampi segmentations. DenseNet achieved an average accuracy of 0.827 on the validation set across different artifact categories, excelling in detecting positioning and banding artifacts, with accuracies of 0.952 and 0.90, respectively. The nnUNet model trained on SF data achieved an average Dice Similarity Coefficient (DSC) of 71% with the SF images-an improvement from 61% DSC using the same nnUNet; similar and significant improvements were obtained for HD, relative volume error, and average symmetric surface distance demonstrating the effectiveness of SF images in improving hippocampal segmentation accuracy. These results indicate that our automated methods effectively improve image quality assessment and hippocampal segmentation in uLF MRI, supporting the mission of the LISA 2024 Challenge and the potential adoption of portable MRI systems in resource-limited regions.

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

Quality Assurance and Hippocampal Segmentation on Low-Field Pediatric Magnetic Resonance Images

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年

查看全文
  • Austin Tapp,
  • Rahimeh Rouhi,
  • Jeffrey Tanedo,
  • Shreyash Zanjal,
  • Sean Deoni,
  • Marius George Linguraru,
  • Natasha Lepore

摘要

Portable ultra-low-field (uLF, i.e., 0.064T) magnetic resonance imaging (MRI) offers a solution to scarce radiological alternatives of resource-limited regions; however, in such regions, MRI system operators and radiologists are novices to the underrepresented modality. Therefore, automatic methods that confirm image acquisition of appropriate quality for diagnosis and segment and measure critical anatomical structures are required to support rural sites. This paper describes our approach to two tasks presented in the LISA 2024 Challenge: (1) quality assurance of pediatric low-field MRI data using DenseNet, and (2) segmentation of the bilateral hippocampi using nnUNet. As uLF MRI natively introduces unique image quality challenges, we trained seven DenseNet264 models, each designed to detect a specific artifact type: noise, zipper, positioning, banding, motion, contrast, and distortion. Similarly, as uLF MRI struggles to offer strong anatomical delineation, we enhanced challenge provided 0.064T low-field MRI scans using a custom Super-Field Network (SFNet) to generate high-quality Super-Field (SF) images, then trained an nnUNet model on these SF images with the LISA Challenge provided hippocampi segmentations. DenseNet achieved an average accuracy of 0.827 on the validation set across different artifact categories, excelling in detecting positioning and banding artifacts, with accuracies of 0.952 and 0.90, respectively. The nnUNet model trained on SF data achieved an average Dice Similarity Coefficient (DSC) of 71% with the SF images-an improvement from 61% DSC using the same nnUNet; similar and significant improvements were obtained for HD, relative volume error, and average symmetric surface distance demonstrating the effectiveness of SF images in improving hippocampal segmentation accuracy. These results indicate that our automated methods effectively improve image quality assessment and hippocampal segmentation in uLF MRI, supporting the mission of the LISA 2024 Challenge and the potential adoption of portable MRI systems in resource-limited regions.