Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Originality: CDDNwithTDC is new and novel: its cascading dense blocks with dilated convolution and its two-step data consistency layer. The author adequately cites relevant works. Quality/Clarity: This is a complete piece of work. The authors are honest about the proposed solution's strengths and weaknesses. The solution may take longer to train and test than some of the prior arts. Nonetheless, the solution uses much fewer number of parameters while setting the new state of the art as far as its performance (PSNR and SSIM) is concerned. The authors also provide an example image with the lesion region zoomed in so that the reader can qualitatively compare the effect of the extra data consistency step in the proposed model. The experiments are also thorough. It showed the effect of the sampling rate and the number of parameter on the model performance (PSNR). The mathematical formulation is sound and clear as well. Overall, it was pleasant reading and following the paper. Significance: The proposed solution and the results reported in the paper are very significant from both algorithmic and application perspectives.
The authors have done a commendable job at developing a model to reconstruct MRI from k-space and evaluating it quantitatively. I'm raising the following possible questions/concerns: - Since the test set is small (3 patients), I would have liked to see k-fold cross validation. - A fair test to compare the reconstructed image to the ground truth could be radiologists' interpretation of them. Quantitative and anecdotal qualitative assessment is great, but it is also important to know if the radiologists' interpretation would have been consistent between the reconstructed and ground truth images. - Have the authors looked into using the FastMRI dataset? I do not have a clinical background, but I wonder if the FastMRI dataset could have been used as a second dataset for reproducibility and generalization of the proposed approach.
I found this paper hard to read. Although I am familiar with both the physics of MRI and deep learning (and have worked at the intersection of the two), it was difficult to follow the exact details of the optimization objective, the data consistency term, etc. More generally, this paper seems to be only a marginal advance over other methods. It seems to fall into the category of "Here is a new DNN architecture with features tuned to a specific problem, and some experiments showing that it works." There is no theoretical justification for the choices made, although they seem reasonable given the context of the problem statement -- however, the final experiments show that the method performs only slightly better than competing DNN-based reconstruction methods, and given the error bars show, it is not clear that the results are statistically significant. Given the readability issues; the ad hoc nature of the architecture; and the less-than-stellar results, I do not believe that this paper crosses the bar for publication in NeurIPS.