Equivariant spatio-hemispherical networks for diffusion MRI deconvolution
Authors: Axel Elaldi, Guido Gerig, Neel Dey
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, we achieve state-of-the-art deconvolution results on two widely used simulated d MRI benchmarks with known ground truth. On real in vivo human d MRI, our method yields more spatially coherent f ODF fields and higher robustness to changes in resolution from research-grade to clinical standards of single-shell low-angular protocols. Lastly, as the achieved efficiency gains enable training on a large set of human datasets, we can now train a single network for amortized inference on new human d MRI, as opposed to the subject-specific optimization of RT-ESD.4 Experiments We first quantify the runtime and memory efficiency gains produced by our contributions. We then analyze their use across a diversity of d MRI deconvolution settings on both real and synthetic datasets. |
| Researcher Affiliation | Academia | Axel Elaldi New York University axel.elaldi@nyu.edu Guido Gerig New York University gerig@nyu.edu Neel Dey MIT CSAIL dey@csail.mit.edu |
| Pseudocode | No | The paper describes its methods through text and diagrams, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/Axel Elaldi/fast-equivariant-deconv. |
| Open Datasets | Yes | We leverage three public d MRI datasets for validation. Di SCo [54] is a synthetic d MRI dataset with three volumes, split into training, validation, and testing volumes, with high-angular resolution sampling (4 shells each with 90 gradients) provided with ground truth f ODF. We then use 100 in-vivo unrelated d MRI scans from the HCP young adult dataset [70]... The last dataset is Tractometer [49]... |
| Dataset Splits | Yes | Di SCo [54] is a synthetic d MRI dataset with three volumes, split into training, validation, and testing volumes...We then use 100 in-vivo unrelated d MRI scans from the HCP young adult dataset [70], split into 65 training, 15 validation, and 20 testing volumes... |
| Hardware Specification | Yes | All comparisons are performed using an Nvidia RTX-8000 GPU.All deep network experiments were performed using a single RTX8000 GPU and used less than 16GB of system RAM. |
| Software Dependencies | No | The paper mentions software tools like MRtrix, FSL, and Free Surfer, but does not specify their version numbers or any other software dependencies with specific versions. |
| Experiment Setup | Yes | Our training regimen uses a batch size of 8 and trains for 50 epochs with an initial learning rate of 3 10 3, which is halved after the 25, 35, and 45 epochs. The optimized objective function encompasses a combined Dice and Cross-Entropy loss with class-dependent weight to address imbalanced labels.All models are trained for 50 epochs with a batch size of 16 patches using the Adam optimizer [41] with an initial learning rate set at 1.7 10 2. The learning rate is decayed by a factor of ten after the 30, 40, and 45 epochs. For unsupervised models, we tuned the regularization weights using the ESD model on the Di SCo validation volume, and further tune λtv for SHD: λnn = 10 1, λsparse = 5 10 5, and λtv = 5 10 1. |