MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI

Area of Science:

• Medical Imaging
• Machine Learning
• Musculoskeletal (MSK) Imaging

Background:

• Deep learning is crucial for MRI applications like reconstruction and artifact removal.
• Current datasets primarily focus on brain and knee imaging, limiting model reliability across diverse anatomies.
• There's a need for comprehensive datasets to study deep learning model performance in varied anatomical contexts.

Purpose of the Study:

• Introduce MosaicMRI, the largest open-source raw musculoskeletal MRI dataset.
• Facilitate training and evaluation of machine learning methods for MSK imaging.
• Investigate the impact of dataset size and anatomical diversity on deep learning model performance.

Main Methods:

• Compiled a diverse dataset of 2,671 fully sampled raw MSK MRI volumes (80,156 slices).
• Included variations in volume orientation, imaging contrasts, anatomies (spine, knee, hip, etc.), and coil numbers.
• Utilized VarNet for accelerated reconstruction and conducted experiments on model scaling and cross-anatomy generalization.

Main Results:

• Models trained on combined anatomies outperformed anatomy-specific models in low-sample regimes.
• Identified specific body part groups (e.g., foot and elbow) that generalize well across each other.
• Demonstrated that performance under domain shifts is influenced by training set size, anatomy, and acquisition protocols.

Conclusions:

• MosaicMRI provides a valuable resource for advancing deep learning in MSK imaging.
• Anatomical diversity in training data enhances model robustness and generalization, particularly in data-limited scenarios.
• Cross-anatomical correlations can be exploited to improve deep learning model performance in MRI.