Abstract
Deep learning models for MRI reconstruction often suffer from a "specialization" problem, where networks trained on specific anatomies fail to generalize to others. This paper introduces a universal framework that leverages Anatomy-SPecific Instance Normalization (ASPIN) and knowledge distillation to achieve high-quality reconstruction across multiple body parts. The proposed model not only outperforms specialized "expert" models but also demonstrates extreme parameter efficiency, requiring the training of less than 0.1% of parameters to adapt to new, small-scale datasets.
Introduction
Magnetic Resonance Imaging (MRI) is a critical diagnostic tool, yet its sequential data collection in "k-space" makes it inherently slow. To accelerate the process, clinicians often employ "undersampling," which frequently introduces blurriness and artifacts. While Deep Learning (DL) has emerged as a powerful solution for image restoration, current models are typically anatomy-specific. A network optimized for brain scans often performs poorly on knee scans due to statistical shift—the differing data distributions across various body parts. Naive attempts to train a single "shared" model on mixed datasets often result in degraded performance, creating a "jack of all trades, master of none" scenario. Furthermore, retraining large models for new clinical applications is computationally prohibitive, especially when only small datasets are available.
Methods
The authors propose a universal framework based on a standard cascaded network (D5C5), enhanced by two key innovations:
- Anatomy-SPecific Instance Normalization (ASPIN): Instead of standard normalization, the framework uses parallel ASPIN modules. Each anatomy (e.g., brain, knee) is assigned its own set of learnable parameters ($\gamma$ and $\beta$). This allows the network to adapt to the specific statistical distribution of a body part while sharing universal convolutional features across all anatomies.
- Model Distillation: To ensure the universal model matches or exceeds the performance of specialized ones, the authors employ "knowledge distillation." Independent "expert" models are first trained for specific anatomies, and the universal model is then trained to mimic their spatial attention maps.
The training follows a four-step pipeline:
- Step (S1): Training independent expert models ($f_1, f_2$).
- Step (S2): Pre-training the universal model ($f_u$) with ASPIN modules.
- Step (S3): Fine-tuning the universal model using distillation from the experts.
- Step (S4): Generalizing to new anatomies by inserting a new ASPIN module and training only those parameters while freezing the rest of the network.
Datasets and Experimental Setup
The study validated the framework using five public datasets:
- Base Training (Large-scale): NYU fastMRI database, comprising 310,798 brain slices and 274,079 knee slices.
- Generalization Testing (Small-scale): ACDC (cardiac, 25,251 slices), CHAOS (abdomen, 1,268 slices), and Medical Segmentation Decathlon (prostate, 904 slices).
Results and Discussion
The universal model was evaluated at 4x and 6x acceleration rates, yielding several key insights:
- Superior Performance: On brain scans at 4x acceleration, the universal model achieved a Peak Signal-to-Noise Ratio (PSNR) of 43.16 dB, significantly outperforming both the independent expert (40.85 dB) and the naive shared model (40.78 dB). This suggests that seeing multiple anatomies helps the network learn more robust, generalized features of high-quality MRI.
- Parameter Efficiency: Adapting the model to a new anatomy, such as the prostate, requires training only 1,280 new parameters. This is less than 0.1% of the total model parameters, making it highly suitable for clinical deployment where data and compute may be limited.
- Small Data Robustness: Even with the limited prostate dataset (~900 slices), the ASPIN-adapted universal model outperformed a model trained from scratch on the same data.
Conclusion
This work demonstrates that a universal MRI reconstruction model is not only feasible but can exceed the performance of specialized experts. By decoupling anatomy-specific statistics via ASPIN from universal feature learning, the framework provides a scalable and efficient path for clinical MRI acceleration. Future research will focus on extending this approach to multi-coil arrays and complex-valued data to better align with real-world clinical workflows.
Summary updated