Development of high-resolution hemodynamic biomarker calculation method for cerebral aneurysms using 4D-Flow and deep learning

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K09175
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 56010:Neurosurgery-related
• Research Institution: Seirei Christopher University (2023); Nagoya University (2021-2022)
• Principal Investigator: Isoda Haruo 聖隷クリストファー大学, リハビリテーション科学研究科, 臨床教授 (40223060)
• Co-Investigator(Kenkyū-buntansha): 平野 祥之 名古屋大学, 医学系研究科(保健), 准教授 (00423129)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: artificial intelligence / deep learning / noise reduction / 4D Flow MRI / intracranial aneurysm / fluid dynamics / hemodynamics

Outline of Final Research Achievements

Teaching data were prepared using the three dimensional and three directional velocity data in cerebral arteries from computational fluid dynamics (CFD) results based on magnetic resonance fluid dynamics (MRFD) results of cerebral arteries containing aneurysms as ground truth data and the data obtained by adding noise that mimics MRFD noise to these CFD data as input data. Prioritising accuracy improvement through noise reduction over higher resolution, the aforementioned teaching data were trained on the Win5-RB model, and deep learning models for noise reduction were developed for 'spatial 2D', 'spatial 2D + time', 'spatial 3D' and 'spatial 3D + time'. The degree of noise reduction was evaluated using the angular similarity index of the velocity vectors, magnitude similarity index of the velocity vectors, velocity vector fields, streamlines and wall shear stress, with the 'spatial 2D + time' noise reduction model being the best.

Free Research Field

Neuroradiology

Academic Significance and Societal Importance of the Research Achievements

脳動脈瘤の発生・成長・破裂に脳動脈瘤の血流動態（壁剪断応力など）が大きな役割を担っており、脳動脈瘤のリスクを予測するバイオマーカーになり得る。ヒトの血流解析には磁気共鳴流体解析(MRFD)と計算流体解析(CFD)がある。MRFDはヒトから直接データを収集できる利点はあるが、ノイズによる精度低下・低空間分解能・低時間分解能の欠点がある。一方、CFDは高時間分解能・高空間分解能・高精度であるが、処理時間や計算時間（3時間から1日）が掛かる欠点があり、臨床応用しにくい。本研究のように、深層学習モデルでMRFDデータのノイズ除去を行い、CFDと同等の精度が即座に得られれば、臨床現場で有用と考えられる。