Hyperparameter Selection for Classic Reconstruction Algorithms

[MatthiasLen]

Problem

Current selection of hyperparameters for classical reconstruction (recon) algorithms is ad-hoc and primarily based on visual inspection of FastMRI knee singlecoil val/test sets. This limits performance.

Objective

Contuct a systematic search for:

• Identifying key hyperparameters in all classic recon algorithms implemented in the codebase
• Determining meaningful value ranges or grid/search spaces for these hyperparameters
• Benchmarking/evaluation strategy: define a few metrics and maybe a small dataset for proper hyperparameter validation (beyond visual inspection)
• Automating or at least documenting the process for hyperparameter selection

Acceptance Criteria

• Survey of all existing classical recon algorithms in the repository with their current default hyperparameters
• Documentation or script for systematic search or validation (e.g., grid search, cross-validation, or other method)
• Updated code or configuration to support reproducible hyperparameter setting
• Results or guidance on selected hyperparameters per algorithm

[MatthiasLen]

@melanie-dohmen-bayer maybe confirm with Yuning before you run the experiements. I set the initial parameters "just that it works" - hence there is probably the need for adjustment in order to configure the recon algos in a fair and more realistic way.

Maybe try some different settigs yourself or look in the liuterature what others use.