Background: Magnetic Resonance Imaging (MRI) is an imaging technique used in daily practice to form diagnostic images of different organs of the body. Its unique features such as superior soft-tissue contrast, elimination of ionizing radiation, and accurate response to functional changes make it an exceptionally well-accepted tool to aid clinical diagnosis. However, due to the underlying physics that governs the generation of images with MRI, its imaging speed is relatively slow compared with other widely-used imaging modalities (e.g. CT and ultrasound), making it uncomfortable sometimes for patients. Therefore, it is desirable to accelerate the image acquisition speed with MRI but without sacrificing image quality.
Goal: To develop a deep learning based image reconstruction method that can recover high-resolution MR images from low-resolution images acquired with accelerated MRI.
Method: A). Developing a supervised deep learning method to define the non-linear mapping for low-resolution and high-resolution image pairs. To make better use of the multi-coil data acquired with MRI and separate the method from conventional natural image super-resolution task, a data consistency layer is incorporated into the convolutional neural network (CNN). This enables the recovering of fine structures on MR images, which is crucial for the detection of the abnormal regions, especially when they are small. B). Validation of the proposed deep learning method on the different field strength of MR scanner (1.5 Tesla and 0.35 Tesla) and different application (abdominal imaging and cardiac imaging) to check the sensitivity of the network to different imaging environment and organs to be imaged. C). Comparing the images generated by the deep learning algorithm with images generated by commercially available method with moderate acceleration factor on the scanner.
Result: This proposed deep learning based reconstruction method is able to generate artifact-free, high-resolution image acquired at different field strength and organ of interest. Image acquisition time can be reduced up to 6-fold without compromising image quality. This directly implies that a conventional 5 minutes scan can be shortened to less than 1 minute. Compared with commercially available algorithms, the proposed method can achieve greater time reduction and at the same time providing diagnostically useful images.