Study of the quantitative evaluation factors for a deep learning-based improved magnetic resonance imaging

Our research focused on the feasibility of improving the low-field T2 magnetic resonance (MR) images in terms of using a complex and in-depth learning-related algorithm. Set of unpaired images (T2-weighted 0.06 T MR image and 1.5 T MR image for separate individuals) were utilized in two clinical trials/sequences. The trials were conducted to identify deformations in a 1.5 T MR image according to the 0.06 T MR image to fit the image and size of the unpaired set. Afterwards, the cyclic-generative adversarial network (GAN) was applied to produce an artificial MR image of the 0.06 T MR image with reference to the original or the deformed 1.5 T MR image. In the end, an improved 0.06 T MR image was produced using the traditional GAN supplemented by the artificial MR image. T2 and flair MR images were verified for matching T1 and T2 testing models in the context of the traditional GAN. The outcomes relating to the optimized trial based on the improved MR image indicated a measurable improvement of the signal with a positive relationship between the original and the improved images. Quantitative variables were applied to assess the quality of the images, along with other settings, such as the signal improvement ratio and the signal-to-noise ratio (SNR), as well as the variable between the original and the improved MR images. The combination of assessment variables demonstrated that T2 photographs were better in terms of the T1 testing model compared to the T2 testing model.