2023 Volume 41 Issue 2 Pages 78-87
Parallel imaging (PI), a fast-scanning technique, has been developed to shorten the scan time of magnetic resonance imaging. However, this technique generates a wide range of noise level and distributes them non-uniformly in the reconstructed image, which requires appropriate noise reduction according to the region and noise level of the image. In this study, we propose a method to adaptively improve the image quality according to the noise level of the image using multiple trained convolutional neural networks (CNNs) even when the training data are scarce. In our method, the training data are separated into multiple datasets based on the g-factor map, which indicates the features of the noise distribution in the imaging region, and a CNN is trained on each dataset. We defined Blur as an index of the low definition of a fine structure and evaluated the performance of the proposed method using PI images captured at three-times high speed as an input. As a result, we have confirmed that Blur is lower than that of a single CNN, and the signal-to-noise ratio exceeds +70 ％, which is equivalent to that of a full-sampled image.
