Week 4 - Alexandra Roberts

­­­­­In the previous post, I described the creation of a training dataset for learning the mapping between 7T analogs and 3T MRI datasets. One surprising nuance a clinician shared with me this week is that a super-resolution approach to 3T is in some cases preferable to an actual 7T scan. While the higher magnetic field strength results in increased spatial resolution, it also results in additional inhomogeneities. There are practical considerations as well, a 7T magnet is significantly heavier than a 3T magnet and can require structural reinforcements to be installed, as the weight of a 7T system is around 25 tons, about double the weight of its 3T counterpart.

Several approaches exist for mapping 3T to 7T analogs. One approach is the continuous representation of an image, allowing super-resolution by an arbitrary scale factor. The continuous representation has an important advantage in that super-resolution methods are often scale-specific. This means that the mapping learned by a neural network will only be effective over scales included in the training data. For example, if a training data set includes low-resolution images downsampled by a factor of 2 and high-resolution counterparts, its performance on a low-resolution test image that is downsampled by a factor of 3 is not guaranteed. Since magnets are manufactured with discrete strengths (0.5 T, 1 T, 1.5 T, 3T, 5T, 7T,…), this limitation is less pronounced than in other settings. However, in the context of MRI, approaching continuous representations adds value beyond super-resolution, such as reducing error based on data discretization.

One network architecture known as a Very Deep Super Resolution (VDSR) generates high-resolution images by calculating the residual image between a traditionally upsampled low-resolution and high-resolution image. A residual image can be calculated for each pair of low and high-resolution images in the training set, and this informs the mapping between the upsampled (by some traditional method such as bicubic interpolation) low-resolution image and its generated super-resolution counterpart.

I divided the dataset into training, validation, and testing subsets. Several cases were set aside to create the testing subset, which is unseen by the network during training. Here is the result of the VDSR for a low-resolution test case, zoomed in by a factor of 5 to demonstrate the super-resolution impact. Note that the images have been brightened uniformly for clarity. The VDSR 7T analog is a super-resolved version of the 3T analog and shows improvement at the vessel boundaries.