Split-slice training and hyperparameter tuning of RAKI networks for simultaneous multi-slice reconstruction.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Reconstruction

Background:

• Simultaneous multi-slice (SMS) acquisitions are crucial for high temporal resolution functional MRI and high-resolution diffusion MRI.
• Deep learning (DL) reconstruction methods, like robust artificial neural networks for k-space interpolation (RAKI), are emerging for unaliasing accelerated SMS data.

Purpose of the Study:

• To systematically evaluate the impact of hyperparameter choices on RAKI network performance for SMS neuroimaging.
• To introduce a novel training data generation technique using a split-slice formalism for RAKI networks.

Main Methods:

• RAKI networks were trained with varied hyperparameters and with/without split-slice data augmentation.
• Networks were tested on five diverse datasets, including Human Connectome Project harmonized data.
• Unaliasing performance was quantified using L1 errors against calibration data.

Main Results:

• Split-slice training substantially improved RAKI network performance across most hyperparameter settings.
• Optimal unaliasing was achieved with 3-layer RAKI networks, specific filter counts, batch normalization, and no dropout, using split-slice data.
• Networks trained without split-slice augmentation exhibited signs of overfitting.

Conclusions:

• Split-slice training is a key factor in enhancing the performance of RAKI networks for SMS neuroimaging.
• Further performance gains in unaliasing can be realized through meticulous hyperparameter tuning of these DL reconstruction networks.