MR fingerprinting Deep RecOnstruction NEtwork (DRONE) (1710.05267v3)

Abstract: PURPOSE: Demonstrate a novel fast method for reconstruction of multi-dimensional MR Fingerprinting (MRF) data using Deep Learning methods. METHODS: A neural network (NN) is defined using the TensorFlow framework and trained on simulated MRF data computed using the Bloch equations. The accuracy of the NN reconstruction of noisy data is compared to conventional MRF template matching as a function of training data size, and quantified in a both simulated numerical brain phantom data and acquired data from the ISMRM/NIST phantom. The utility of the method is demonstrated in a healthy subject in vivo at 1.5 T. RESULTS: Network training required 10 minutes and once trained, data reconstruction required approximately 10 ms. Reconstruction of simulated brain data using the NN resulted in a root-mean-square error (RMSE) of 3.5 ms for T1 and 7.8 ms for T2. The RMSE for the NN trained on sparse dictionaries was approximately 6 fold lower for T1 and 2 fold lower for T2 than conventional MRF dot-product dictionary matching on the same dictionaries. Phantom measurements yielded good agreement (R2=0.99) between the T1 and T2 estimated by the NN and reference values from the ISMRM/NIST phantom. CONCLUSION: Reconstruction of MRF data with a NN is accurate, 300 fold faster and more robust to noise and undersampling than conventional MRF dictionary matching.

• The paper demonstrates a deep learning method that replaces large dictionaries, cutting storage to 5% and accelerating reconstruction by 300 times.
• The neural network, implemented in TensorFlow with ADAM optimization, achieves RMSE values of 3.5 ms for T1 and 7.8 ms for T2, ensuring superior accuracy.
• The DRONE approach enhances robustness through Gaussian noise training and supports flexible adaptation to varied MR pulse sequences.

An Analytical Overview of the MR Fingerprinting Deep Reconstruction Network (DRONE)

Introduction

The paper introduces a novel approach for the rapid and efficient reconstruction of multi-dimensional Magnetic Resonance Fingerprinting (MRF) data utilizing Deep Learning methodologies, specifically through a neural network (NN) referred to as the Deep Reconstruction Network (DRONE). This approach addresses the inherent complexities and computational burdens associated with traditional MRF reconstruction methods that rely heavily on extensive dictionaries, by leveraging the efficiency and compact nature of deep learning algorithms.

Methodology

The proposed method involves developing a fully connected neural network with an input layer corresponding to the 25-point trajectory of magnitude images acquired using an optimized Echo-Planar Imaging (EPI) MRF sequence. The NN, implemented in the TensorFlow framework, comprises four layers, including two hidden layers each equipped with 300 neurons, and utilizes the ADAM stochastic gradient descent optimization. Training is executed on simulated MRF data generated via Bloch equations, with an approximate convergence time of 10 minutes, significantly expediting the traditional T1 and T2 mapping tasks.

The neural network is trained with Gaussian noise-augmented data to enhance robustness, which is crucial for handling real-world noisy signal inputs. The training dictionary consists of approximately 79,900 entries, underscoring the capability of the network to process a significant volume of data efficiently. The resultant neural network offers a remarkable reduction in storage requirements, utilizing only 5% of the space needed for conventional dictionary storage.

Results

The performance of the DRONE model is rigorously validated using both numerical simulations and empirical data from the ISMRM/NIST phantom, demonstrating superior accuracy and robustness. The root-mean-square error (RMSE) for the NN in reconstructing simulated brain data stands at a compelling 3.5 ms for T1 and 7.8 ms for T2, which is substantially lower compared to traditional methods. The network shows excellent agreement (R = 0.99) with reference values from phantom measurements, illustrating its effectiveness.

In practical terms, the NN reconstruction is 300 times faster than the conventional MRF template matching techniques due to its rapid feedforward processing. This acceleration is pivotal in advancing the clinical viability of MRF by mitigating the otherwise prohibitive computational demands.

Discussion

The compact architecture of the DRONE network addresses the principal challenges of conventional MRF methods, which involve exponential growth in dictionary size with high-dimensional applications. By providing a continuous functional mapping of signal magnitudes to tissue parameter values, the NN overcomes the discrete limitations of dictionary matching, offering improved noise resilience and overall reconstruction accuracy.

Furthermore, the authors highlight the flexibility of the model with respect to pulse sequence variations, signaling potential adaptability across diverse MR imaging applications. The scalability of the DRONE model for more complex reconstructions with additional tissue parameters, however, may require deeper network architectures and more advanced training methodologies, which the paper proposes to explore in future investigations.

Conclusion

Overall, the DRONE approach exemplifies the integration of Deep Learning into MRF reconstruction, achieving an impressive balance between speed, accuracy, and computational efficiency. This paper contributes substantially to the field of magnetic resonance imaging by demonstrating the potential of neural networks to surmount issues associated with traditional MRF methods, paving the way for future advancements in the utilization of machine learning for medical imaging applications. Future developments could further enhance the capabilities of DRONE, incorporating deeper networks and exploring varied activation functions to optimize performance across a wider array of imaging contexts.