Spatio-Angular Convolutions for Super-resolution in Diffusion MRI

Authors: Matthew Lyon, Paul Armitage, Mauricio A Álvarez

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the PCCNN performs competitively while using significantly fewer parameters. Moreover, we show that this formulation generalises well to clinically relevant downstream analyses such as fixel-based analysis, and neurite orientation dispersion and density imaging. ... 4 Experiments and Results
Researcher Affiliation Academia Matthew Lyon University of Manchester matthew.s.lyon@.manchester.ac.uk Paul Armitage University of Sheffield p.armitage@sheffield.ac.uk Mauricio A Álvarez University of Manchester mauricio.alvarezlopez@manchester.ac.uk
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code for this work is available at github.com/m-lyon/dmri-pcconv.
Open Datasets Yes d MRI data from the WU-Minn Human Connectome Project (HCP) [38] were used for training, validation and testing.
Dataset Splits Yes d MRI data from the WU-Minn Human Connectome Project (HCP) [38] were used for training, validation and testing. ... Models were trained on twenty-seven subjects from the HCP dataset, while three subjects were used for validation during development.
Hardware Specification Yes Models were trained using 4 NVIDIA A100 s with a batch size of 16, and an ℓ1 loss function, for 200,000 iterations using Adam W [23].
Software Dependencies No The paper mentions software like 'Ray Tune [20]', 'MRtrix3 [37]', and 'cu DIMOT [17]' but does not provide specific version numbers for them.
Experiment Setup Yes Models were trained using 4 NVIDIA A100 s with a batch size of 16, and an ℓ1 loss function, for 200,000 iterations using Adam W [23]. ... Hyperparameters for the PCCNN were selected through a random grid search with Ray Tune [20]. ... Each PCConv layer or residual PCConv block is followed by a rectified linear unit (Re LU), excluding the final layer. ... Each hypernetwork is composed of two dense layers, each followed by a leaky Re LU with a negative slope of 0.1, and a final dense layer with output size 1 and no subsequent activation. ... The input angular dimension size was determined via qin U(qsample), qsample = {6, 7, ..., 19, 20}.