Improving High-Frequency Details in Cerebellum for Brain MRI Super-Resolution

Hanzhi Wang, David Marshall, DEREK KENTON JONES, Yuhua Li

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Cerebellum, Data-balancing, MRI, Super-Resolution

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: A simple yet effective method for improving high-frequency details in cerebellum for MRI super-resolution.

Abstract:

Deep-learning-based single image super-resolution (SISR) has attracted growing interest in clinical diagnosis, especially in the brain MR imaging field.
Conventionally, SISR models are trained using paired low-resolution (LR) and high-resolution (HR) images, and image patches rather than the whole images are fed into the model to prevent hardware memory issues.
However, since different brain regions have disparate structures and their size varies, such as the cerebrum and the cerebellum, models trained using image patches could be dominated by the structures of the larger region in the brain and ignore the fine-grained details in smaller areas.
In this paper, we first investigate the capacities of several renowned models, by using more blurry LR images than previous studies, as input.
Then, we propose a simple yet effective method for the conventional patch-based training strategy by balancing the proportion of patches containing high-frequency details, which makes the model focus more on high-frequency information in tiny regions, especially for the cerebellum.
Our method does not depend on model architectures and this paper focuses solely on the T1-weighted brain MR images.
Compared with the conventional patch-based training strategy, the resultant super-resolved image from our approach achieves comparable image quality for the whole brain, whereas improves significantly on the high-frequency details in the cerebellum.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2882