MRI Banding Removal via Adversarial Training
Authors: Aaron Defazio, Tullie Murrell, Michael Recht
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We report the results of a blind comparison against a strong baseline by a group of expert evaluators (board-certified radiologists), where our approach is ranked superior at banding removal with no statistically significant loss of detail. Our paper is structured as follows: In Section 2 we formally describe the MRI reconstruction problem as it applies to current clinical MRI scanners. In Section 3 we describe how to augment standard deep-learning based MRI reconstruction methods with our orientation adversary. Section 4 describes the state-of-the-art reconstruction model that we use in our experiments. In Section 5 we describe the masking procedure we use, in Section 6 we describe the classical baseline that we compare against, and in Section 7 we detail how our model was trained. Finally in Section 8 we detail the results of a blind evaluation by radiologists. |
| Researcher Affiliation | Collaboration | Aaron Defazio Tullie Murrell Facebook AI Research, Facebook New York Michael P. Recht Department of Radiology NYU Grossman School of Medicine |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Full source code is available in the supplementary material. |
| Open Datasets | Yes | We trained our models on the knee scans from the fast MRI dataset [22]. [22] Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J. Muckley, Mary Bruno, Aaron Defazio, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdzal, Adriana Romero, Michael Rabbat, Pascal Vincent, James Pinkerton, Duo Wang, Nafissa Yakubova, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, and Yvonne W. Lui. fast MRI: An open dataset and benchmarks for accelerated MRI. 2018. |
| Dataset Splits | Yes | We trained with a factor of 4 acceleration, using 16 central k-space lines, using the train/test splits distributed as part of the fast MRI dataset. Each of the six board-certified radiologists in our study were given a set of 20 of 40 volumes from the validation set (Each volume is approximately 25 slice images), so that each volume of the 40 was evaluated three times independently. |
| Hardware Specification | No | The paper mentions '8-gpu system' but does not provide specific details about the GPU models (e.g., NVIDIA A100, RTX 2080 Ti) or other hardware components like CPU or memory. This level of detail is not sufficient for full reproducibility as per the strict requirements. |
| Software Dependencies | No | The paper does not specify version numbers for any software dependencies, such as Python, PyTorch, or other libraries used. |
| Experiment Setup | Yes | Training consisted of 100 pre-training epochs, where the adversarial term is not used (but the flip operator is included), using ADAM with learning rate 0.0003, momentum 0.9 and batch-size 8 (1 per GPU on an 8-gpu system) and no weight decay. The pre-training was followed by 60 epochs of training including the adversarial term, with learning rate 0.0001, and adversary gradient regularization strength γ = 0.1. |