MRI Compressed Sensing Image Reconstruction Using Deep Learning

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K04423
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 21030:Measurement engineering-related
• Research Institution: Utsunomiya University
• Principal Investigator: ITO SATOSHI 宇都宮大学, 工学部, 教授 (80261816)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 磁気共鳴映像法 / 深層学習 / 圧縮センシング

Outline of Final Research Achievements

Deep learning-based image reconstruction methods produced higher quality images than conventional iterative reconstruction methods in compressed sensing MRI. There was no appearance of unnaturalness in the images, and the reconstruction time was greatly reduced to within 0.5 seconds, which solved most of the initial issues. With respect to learning methods, signal-to-image learning is robust to signal under-sampling patterns, and image-to-image learning is robust to signal under-sampling. Deep learning-based reconstruction is not just a replacement for iterative reconstruction methods, but can relax the conditions for compressed sensing, suggesting the possibility of changing even the measurement method.

Free Research Field

医用画像工学

Academic Significance and Societal Importance of the Research Achievements

MRIの撮像時間を短縮する目的で圧縮センシングを導入する研究を行った．課題であった画質と再構成時間の短縮は高いレベルで満足できる結果を得た．また，深層学習再構成の方式によっても画質や再構成時間，必要な画像データ数などが異なり，それぞれ特徴があることが明らかになった．さらに，深層学習の導入は計測方法の常識までも変える可能性を持ち，今後さらに計測と計算機処理の融合により高いレベルでの計測，さらなる撮像時間の短縮化など新たな次元の計測へ展開する方向性が示唆された．このように深層学習に代表されるデータ駆動科学のMRI応用において新たな知見が得られたことから学術的，社会的意義は大きいものがある．