On the generalizability of diffusion MRI signal representations across acquisition parameters, sequences and tissue types: Chronicles of the MEMENTO challenge

Summary

The MEMENTO challenge tests how diffusion MRI signal models generalize to unseen measurements and different encoding schemes. Results show that fitting choices and hyperparameters matter alongside model selection, with double and oscillating encoding proving especially challenging.

Study overview

Eight teams submitted 80 fits to predict withheld measurements from human and mouse diffusion MRI datasets. Most methods predicted single-diffusion-encoding signals well, with weaker performance at extreme diffusion weightings and on double or oscillating encoding. The comparison emphasizes reporting and optimizing fitting procedures as well as choosing a signal model.

Main findings

• Generalization depended on the acquisition scheme and diffusion weighting.
• Fitting choices and hyperparameters substantially affected prediction performance.