Evaluation of the image quality index with MRI motion artifacts on tumor segmentation using deep learning.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiological Technology, Graduate school of Health Science, Juntendo University, 2-1-1, Bunkyo-ku, Tokyo, 113-8421, Japan.
- Department of Radiological Technology, Graduate school of Health Science, Juntendo University, 2-1-1, Bunkyo-ku, Tokyo, 113-8421, Japan. [email protected].
- Department of Radiological Technology, Faculty of Health Science, Juntendo University, Bunkyo-ku, Japan. [email protected].
- Department of Radiation oncology, Faculty of medicine, Juntendo University, Bunkyo-ku, Japan. [email protected].
- Department of Radiological Technology, Faculty of Health Science, Juntendo University, Bunkyo-ku, Japan.
Abstract
Motion artifacts substantially degrade magnetic resonance imaging (MRI) quality and can reduce diagnostic accuracy. Although deep learning-based segmentation has achieved high performance, the degree of motion-related degradation that begins to impair tumor contour delineation and the corresponding image quality criteria remains unclear. This study investigated the relationship between image quality and the agreement of segmentation outputs obtained from original and motion-degraded MRI images in deep learning-based brain tumor segmentation and explored tentative reference values for image quality metrics. Tumor contours were manually delineated on fluid-attenuated inversion recovery images from patients with glioma, and a segmentation model was tested. Motion artifact images were then generated through simulation, and the relationship between the generated tumor contours and image quality was analyzed. The tumor contours were evaluated using the Dice similarity coefficient (DSC), while image quality was assessed using mean absolute error (MAE), root mean squared error (RMSE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). No strong correlations were observed between DSC and the evaluated image quality metrics. However, SSIM showed a positive association with visual assessment, suggesting that it better reflects tumor boundary visibility and perceived image degradation than other indices. In particular, an SSIM of 0.8 corresponded to a visual assessment score of 2 or higher, indicating a quality level at which observers can visually distinguish tumor boundaries. Because these findings may depend on the number of cases and the segmentation model used, further validation with clinical images is required to establish practical image quality criteria.