Prediction of radiation pneumonitis with machine learning using dose-volume and dose-function features

2023 Fiscal Year Final Research Report

• Project/Area Number: 20K16815
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 52040:Radiological sciences-related
• Research Institution: Tohoku University
• Principal Investigator: Katsuta Yoshiyuki 東北大学, 大学病院, 助教 (90848326)
• Project Period (FY): 2020-04-01 – 2024-03-31
• Keywords: 放射線治療 / 肺炎 / 機械学習 / 人工知能

Outline of Final Research Achievements

A predictive model was developed using support vector machine and random forest. Both models were able to build predictive abilities that exceeded existing methods including V20Gy, V5Gy, and mean lung dose. In order to improve predictive ability, we started (1) applying differential histograms and (2) creating a new dose-function feature. The model that added the differential dose-volume-histogram had superior predictive ability compared to the model constructed only from the integral dose-volume-histogram. Next, we created a dose-function specialized for machine learning that avoids multicollinearity. After adding new cases and investigating the benefits of our dose-function features, we obtained predictive performance from the LASSO and support vector machine methods that is expected for clinical use.

Free Research Field

放射線治療

Academic Significance and Societal Importance of the Research Achievements

進行非小細胞肺がんの治療計画では、放射線誘発性肺炎を回避する目的で肺機能画像上に抽出した高機能肺領域の被ばくを選択的に低減する方法が試みられている。治療の安全性や成績は改善されつつあるが、一定数の症例で致死性を伴う肺炎が発生している。また、近年では免疫療法の普及によって治療成績の改善が可能となったが、投与は肺炎がGrade2未満の症例に限定されており、肺炎の抑制は重要である。このように、肺炎発生の回避は進行肺がん治療の安全性と治療効果の双方に恩恵をもたらすと考えられ、早急な手立てが熱望されている。