A basic investigation of texture analysis and deep learning for positron emission tomography

2021 Fiscal Year Final Research Report

• Project/Area Number: 19K17127
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 52040:Radiological sciences-related
• Research Institution: Hokkaido University
• Principal Investigator: Kobayashi Kentaro 北海道大学, 医学研究院, 客員研究員 (70756311)
• Project Period (FY): 2019-04-01 – 2022-03-31
• Keywords: 核医学 / 人工知能 / radiomics / deep learning / PET / FDG

Outline of Final Research Achievements

FDG PET-CT, which visualize the distribution of glucose metabolism in the body, is used in routine medical practice as a tool to visualize malignant tumors. As with other diagnostic imaging, there are great expectations for AI technology in PET, but there is insufficient basic data to apply it to real-world clinical practice. In this study, we focus on radiomics and deep learning. Radiomics is a technology or research field that quantifies the shape and internal heterogeneity of a lesion using pixel value formulas for diagnosis. On the other hand, deep learning is a technology that uses a deep neural network to include the feature design process in machine learning. In the present study, we applied these methods to clinical PET images as well as phantom images using 3D printer that simulated tumor, brain, and breast. We obtained meaningful data for future AI developments.

Free Research Field

放射線科、核医学、画像診断、人工知能

Academic Significance and Societal Importance of the Research Achievements

RadiomicsをPETに用いた研究は散見されるものの、実用化には至っていない。これはradiomicsの有用性が確立していないことが原因と考えられた。多数の特徴量の中から、どのような特徴量が有用であるかを明らかにしていく研究が不足していたため、今回はこれに焦点をあてた。臨床画像ではground truthが得られないことが多いため、ファントムを併用した。同時に、急速に進化するdeep learning(DL)技術をPETに応用していくことは急務と考えられたため、これもあわせて今回の研究テーマとし、DLが解決できる課題とそうでない課題とを区別するための知見を得ることができた。