Development of Deep Convolutional Neural Network for Computer-Aided Diagnosis of Ischemic Heart Disease

2022 Fiscal Year Final Research Report

• Project/Area Number: 20K20233
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 90130:Medical systems-related
• Research Institution: Niigata University of Health and Welfare
• Principal Investigator: Hasegawa Akira 新潟医療福祉大学, 医療技術学部, 講師 (20749999)
• Project Period (FY): 2020-04-01 – 2023-03-31
• Keywords: 虚血性心疾患 / 深層学習 / 冠動脈CT / プラーク / 自動分類 / 領域抽出

Outline of Final Research Achievements

In order to correctly recognize calcified plaques, which are the cause of low positive predictive value in coronary CT, automatic classification with stents with similar findings of calcified plaques was performed using deep learning. The results showed that VGG-23, which has 6 more convolution layers and 1 more full-connect layer than VGG-16, had the highest accuracy in automatic classification by fine tuning, resulting in an accuracy of 98.0%. In the automatic classification of plaques including low-absorption plaques, low-absorption plaques were not correctly recognized. Therefore, low absorption plaque regions were automatically extracted using U-Net, which can automatically extract regions, and the Dice coefficient was 0.91, indicating that the regions were extracted with high accuracy.

Free Research Field

放射線技術学

Academic Significance and Societal Importance of the Research Achievements

本研究の結果は、深層学習による冠動脈CTでの虚血性心疾患の診断支援の要素技術となる。特に冠動脈CTで陽性的中率が下がる原因となる石灰化プラークを深層学習が正しく認識できることが期待でき、陽性的中率の改善が期待される。また、低吸収プラークの領域も正しく抽出することができたため、今後はプラークを自動認識するだけでなく狭窄率も正確に自動分析することが期待される。冠動脈CTにおける虚血性心疾患の診断支援技術の普及により、さらに迅速かつ正確な診断が期待される。