search

simonsf / Hybrid-DL-IR

Hybrid Deep-learning and Iterative Reconstruction Scheme for Medical Imaging Reconstruction
5 stars 1 forks source link

Discussion on quantization results between MoDL and VarNet #2

Open Breeze-Zero opened 11 months ago

Breeze-Zero commented 11 months ago

Thank you very much for your work, let me gain a lot. I have a few questions about this, and I hope you can take a moment to answer them. The noise and artifacts of MODL and VarNet are too obvious under 3x, which is quite different from the usual experimental conclusion. I guess that your data set contains many organs and different acquisition machines, or the sampling method is different, but I am not sure about the specific reason. How do you view this situation?

simonsf commented 11 months ago

Thanks for your attention!

Referring to the MRI reconstruction experiments we have done and the relevant papers we have read, E2E VarNet does have a high sensitivity for different organ imaging, downsampling strategies, and pulse sequence designing. Compared to the public FastMRI dataset, the downsample masks in our data involve higher diversity, and also our data contains more inpulse sequences.

Our experimental results show that E2E VarNet does not perform well enough in some scenarios. Even with the public FastMRI dataset, the experimental results of the paper [https://proceedings.neurips.cc/paper/2021/hash/7d6044e95a16761171b130dcb476a43e-Abstract.html] show that E2E VarNet performs not as expected in abdominal or knee imaging. We have great interest in improving the imaging speed and quality of MRI, looking forward to your suggestions and ideas.