In magnetic resonance imaging (MRI), motion artifacts vary in intensity and direction, which prevents accurate image diagnosis. However, through use of transfer learning and employing gradient-weighted class activation mapping (Grad-CAM) to create a classification model, the characteristics of these classifications can be clarified. Previous studies have indicated that the model exhibited a greater response to more prominent artifacts, and this response distribution may depend on the difference in the phase-encoding direction of the image. The study aim was to create motion-artifact classification models for different phase-encoding direction and use Grad-CAM to evaluate their effects. Using T2-weighted images from 25 head MRI cases, we performed computational simulation to create 9,000 images with artifacts. The data were divided into 70% training, 30% validation, and GoogLeNet was used to perform transfer learning. The test accuracy was 91.5% for the combined phase-encoding direction of the vertical and horizontal models, 96.4% for the vertical model, and 96.7% for the horizontal model. The Grad-CAM results reflected the characteristics of the models, and the combined model did not show any phase-encoding direction effects.
抄録全体を表示