Construction of AI image diagnosis support systems for diagnosing and predicting serious emergency brain diseases

2023 Fiscal Year Final Research Report

• Project/Area Number: 20K16737
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 52040:Radiological sciences-related
• Research Institution: Juntendo University
• Principal Investigator: Junko Kikuta 順天堂大学, 医学部, 助教 (70613389)
• Project Period (FY): 2020-04-01 – 2024-03-31
• Keywords: 人工知能 / 画像診断 / 頭部画像

Outline of Final Research Achievements

Using brain MRI of 69 young subjects from the Human Connectome Project, we constructed a machine learning model that automatically detects perivascular spaces. After making mixed T1-weighted and T2-weighted images to depict the clear perivascular space on FMRIB Software Library version 6.0, the perivascular space was manually extracted. The concordance rate between the U-Net model and manually detected perivascular spaces was evaluated. As a result, the average value of the Dice index for images of only cerebral white matter was 0.493 ± 0.042, and the Dice index of whole brain images was 0.385 ± 0.064. The Dice index of images of only cerebral white matter was significantly higher than that of whole brain images ( p=0.001).

Free Research Field

画像診断、神経放射線領域、人工知能

Academic Significance and Societal Importance of the Research Achievements

研究成果はThe 22nd Asian Oceanian Congress of Radiologyで学会発表を行い、最優秀賞を受賞した。健常若年者の血管周囲腔は小さいため、手動抽出、機械学習モデルでの自動抽出ともに難しく、本研究はT1強調像をT2強調像で除し、血管周囲腔をより明瞭にした画像を用いた点で新規性があると考える。