Evaluation of the image quality index with MRI motion artifacts on tumor segmentation using deep learning

In magnetic resonance imaging (MRI), motion artifacts are a major factor that degrades image quality and diagnostic accuracy. Although deep learning-based high-precision segmentation techniques have emerged of late, the impact of image quality degradation on segmentation accuracy is not sufficiently understood, and clinically acceptable image quality standards are not yet clearly defined. This study aims to quantitatively evaluate the effects of motion artifacts on brain-tumor segmentation using deep learning and clarify clinically acceptable image quality criteria. Fluid-attenuated inversion recovery ( FLAIR ) images of glioma patients were used to manually delineate tumor contours, and a segmentation model was constructed using an AI development support service based on the contour data. Simulated motion artifact images were generated, and the relationship between segmentation accuracy and image quality was analyzed. Segmentation accuracy was evaluated using the dice similarity coefficient (DSC), while image quality was assessed using mean absolute error, root mean squared error, structural similarity index measure (SSIM), and peak signal-to-noise ratio. No strong correlation was observed between DSC and the four image quality indices; however, SSIM differed significantly between cases with DSC ≥ 0.8 and DSC < 0.8. Furthermore, visual evaluation by three experienced radiological technologists revealed a positive correlation between SSIM and perceived image quality; when an SSIM of 0.8 was used as a threshold, the visual evaluations differed significantly. These results suggest that SSIM is an effective indicator of human visual perception of image quality, and an SSIM threshold of 0.8 may represent a clinically acceptable standard for MRI image quality.