Evaluation on the generalization of a learned convolutional neural network for MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models, particularly convolutional neural networks (CNNs), are increasingly used for accelerating magnetic resonance imaging (MRI) through k-space data reconstruction.
• Assessing the robustness and generalization of these trained networks when applied to data dissimilar to their training sets remains a critical challenge.

Purpose of the Study:

• To quantitatively evaluate how variations in image contrast, human anatomy, sampling patterns, undersampling factors, and noise levels impact the generalization performance of a deep cascade of convolutional neural networks (DC-CNN) for MRI reconstruction.
• To understand the factors influencing the reliability of deep learning models in MRI when applied to diverse datasets.

Main Methods:

• A deep cascade of convolutional neural networks (DC-CNN) with a data consistency layer was trained on datasets with varied parameters (contrast, anatomy, sampling, undersampling, noise).
• The trained DC-CNN was tested on datasets with varying degrees of consistency and inconsistency with the training data to assess generalization.
• Performance was evaluated based on reconstruction quality under different data deviation scenarios.

Main Results:

• Reconstruction quality is highly sensitive to sampling pattern, undersampling factor, and noise level, which correlate with signal-to-noise ratio (SNR).
• Image contrast showed a relatively lower impact on reconstruction quality compared to sampling and noise.
• Significant performance degradation was observed when human anatomy in test data deviated substantially from training data, especially for brain images with high anatomical detail.

Conclusions:

• The generalization of DC-CNNs in MRI reconstruction is significantly influenced by sampling strategies, undersampling ratios, and noise levels.
• Anatomical complexity plays a crucial role in model generalizability, with simpler anatomies (chest, knee) showing better performance than complex ones (brain) under data deviation.
• These findings offer empirical insights into deep learning model generalizability in MRI and underscore the potential of transfer learning for reconstructing images from datasets different from those used during initial training.