Benchmarking MRI Reconstruction Neural Networks on Large Public Datasets
Round 1
Reviewer 1 Report
Dear All,
The proposed publication is strong in it's Statistical computation, Data and Results presentation and Discussion. The novelty backed up well enough and comparison to similar studies is supplied.
What concerns the Introduction - it is not of the same high level as the rest of the publication.
E.g., the pediatric group patients or Parkinson disease patients are scanned according to the specific protocols and if necessary - general anaesthesia is applied in these cases. So motion artifacts are not really that important
Motion artifacts can be excluded by means of special software provided with the scanner, as well as specific short time acquisitions can be applied, instead of normal range scan times. All these means can be easily provided by the vendor, basing on financial capabilities of the hospital. Also, older MRI scanners can be upgraded in order to install specific short time acquisitions.
My suggestion - to improve the level of Introduction, to prove the real importance of your research.
Sincerely,
Author Response
We want to thank the reviewer for their review and interesting comments.
While many works do tackle the problem of MRI reconstruction, it is important to always motivate the research.
The reviewer hinted that because anesthesia could be used to prevent motion, or because short acquisition could be used as well as motion correcting software, working on reconstruction time was not that significant.
However, anesthesia is a very heavy process still reducing accessibility to the exam. We will include this comment though to make sure this passage is clear in lines 21-22.
In addition, even during short acquisitions you can have motions, and this becomes a problem especially at ultra high field (like 7T or 11.7T). And a short acquisition time will lead to an increased reconstruction time with the classical methods, i.e. iterative reconstructions.
A short reconstruction time, also allows to plan rapidly complementary acquisitions to clarify the image-based diagnosis. We added that in lines 26-28.
Reviewer 2 Report
This well written paper describes a benchmark for MRI reconstruction based on artificial neural networks and a recently released dataset, fastMRI, and the OASIS data. Results from different networks on this dataset are compared. Reproducible code and the models trained are provided. The main finding is that it is better to perform more iterations between the image and the measurement spaces instead of having a deeper per-space network. It is well structured and easy to follow. The limitations are well addressed. My only suggestion is that the figures should have been clearer. Deep learning has become a useful tool in the image reconstructionfield to reduce image acquisition time while keeping as many details
as possible. The reviewer would like the authors put some efforts on
describing image reconstruction for accelerating for other imaging modalities,
such as DWI. Instead of using a fixed mask for undersampling, the mask itself can be
learned based on deep learning model. At least some preliminary works
have done so recently. Please give some credits. The MRI fingerprinting has been proposed for quantitative imaging for years.
Recently, deep learning has been implemented in this field.
Please also include this trend. The result evaluation metrics used in image reconstruction are
not only including PSNA and SSIM.
Author Response
We want to thank the reviewer for their review and interesting comments.
About the image quality metrics: there are indeed more metrics that could be used to illustrate the performance of the networks, but PSNR and SSIM are the 2 mainly used and in particular they are used in the fastMRI leaderboard (https://fastmri.org/leaderboards). Moreover, NMSE which is also used in the fastMRI leaderboard is a bit more difficult to interpret.
About mask learning: the mask learning trend is indeed strong, and while we mentioned some approaches trying to optimize the sampling pattern we did not mention any approaches learning the masks or even doing the mask learning jointly with reconstruction. We will mention them in the final manuscript, lines 202-206.
About DWI and MRI fingerprinting: we will be sure to mention that deep learning can be also applied to other imaging modalities in particular in MRI. We will try to make the link with the current work, lines 59-62.