Adapting model-based deep learning to multiple acquisition conditions: Ada-MoDL.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Undersampled parallel MRI data acquisition often requires separate models for different settings.
• Existing methods necessitate training and storing multiple networks, increasing complexity and data requirements.

Purpose of the Study:

• To introduce a single model-based deep network for high-quality reconstructions from undersampled parallel MRI data acquired with multiple sequences and settings.
• To develop an adaptive framework that can generalize across various acquisition parameters, including field strengths and contrasts.

Main Methods:

• A single unrolled deep network architecture (Ada-MoDL) is proposed, adapting to different acquisition settings via feature and regularization parameter scaling.
• A multilayer perceptron model derives scaling weights from conditional vectors representing specific acquisition settings.
• The network parameters and convolutional neural network (CNN) weights are jointly trained using diverse multi-setting data.

Main Results:

• The adaptive framework demonstrates consistently improved performance across all tested acquisition conditions compared to training a single model on all data.
• The proposed scheme requires significantly less training data per setting to achieve good performance compared to independently trained networks.

Conclusions:

• The Ada-MoDL framework enables a single model-based unrolled network for multiple MRI acquisition settings.
• This approach reduces the need to train and store multiple networks and decreases the training data required for each setting.