Aging prediction using deep generative models toward the development of preventive medicine

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K18073
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 90130:Medical systems-related
• Research Institution: The University of Tokyo
• Principal Investigator: Shibata Hisaichi 東京大学, 医学部附属病院, 特任助教 (10780067)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 老化予測 / 人体デジタルツイン / 機械学習 / 深層生成モデル / 確率的デノイジング拡散モデル / 不確実性評価 / 頭部MRI / 全脳容積評価

Outline of Final Research Achievements

I have developed a novel deep generative model that can generate images conditioned on multiple past three-dimensional images. The novelty (outcomes) of this research includes: (1) the ability to predict future MRI images representing the human head from a multitude of MRI scans over the past seven years (multi-point prediction), and (2) unlike previous studies which are mostly limited to comparisons at nearly a single future point, we evaluated the model's prediction accuracy by comparing quantities obtained from the predicted images, such as total brain volume, with those defined over multiple points up to approximately seven years into the future (multi-point accuracy validation). Additionally, I have developed a framework capable of predicting numerous possible but different future human bodies (uncertainty evaluation).

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

本研究の成果を応用することで、例えば(i)予測結果を受診者に提示することで、受診者の健康への意識を高められ、予後の向上に貢献できる可能性、(ii)医師の目ではわからないほど早期から疾患を診断できるコンピュータ支援診断ソフトウェアの開発へ繋がる可能性、(iii)加齢に伴う疾患の発生部位や発生時期の予測による予防的治療が実現できる可能性、(iv)既に病気に罹患している者を撮像した医用画像からその過去を予測し、医者が見落としやすい病変群を明らかにできる可能性などを持ち、広く社会へインパクトを与えられると考えられる。