Synopsis

Keywords: Kidney, Quantitative Imaging, Cancer

In this pilot study, we developed and optimized a spatially-constrained convolutional neural network to accelerate the scan time for 2D kidney Magnetic Resonance Fingerprinting (MRF). Our results suggest that an acceleration factor of 3 can be achieved with the proposed method, which shortens the 2D breath-hold MRF scan from 15 sec to 5 sec. In addition, the deep learning based approach can be applied for T₁ and T₂ quantification of both normal renal tissues and pathologies including renal cell carcinoma.

Introduction

Magnetic Resonance Fingerprinting (MRF) is a new quantitative MR imagine technique, which can provide rapid and simultaneous quantification of multiple tissue properties (1). Recently, deep learning based methods have been integrated with MRF to further improve its speed for both 2D and 3D acquisitions (2, 3). However, most of these studies are focused on stationary brain imaging, where its utility in quantitative abdominal imaging has not been widely explored. Application of the MRF method in these organs faces more technical challenges due to respiratory motions (4). In addition, most of the deep learning methods are tested with MRF measurement obtained from normal tissues and its applicability in characterization of abnormal tissues has not been evaluated. The objectives of this pilot study are to 1) develop and optimize a spatially-constrained convolutional neural network (CNN) to accelerate 2D kidney MRF acquisition (5) and 2) evaluate its capability in quantitation of T₁ and T₂ relaxation times for renal cell carcinoma (RCC).

Methods

All imaging was performed on a Siemens 3T MR scanner (Skyra or Vida). The kidney MRF data was acquired using a FISP-based sequence with a FOV of 40×40 cm, matrix size of 256×256 and slice thickness of 5 mm (4). A total of 1728 MRF time frames were acquired in a single breath-hold of ~15 seconds. Kidney MRF dataset was acquired from 45 subjects. These included 22 normal volunteers and 23 patients with a variety of pathologies such as RCC, autosomal recessive polycystic kidney disease (ARPKD), and cystic fibrosis (CF). An average of 3~5 slices were acquired from each subject, yielding a total of 147 slices from all the subjects for the training and testing purposes.
A UNet-based, spatially constrained CNN with residual channel attention was adopted and optimized to retrospectively accelerate 2D kidney MRF acquisition (5). The original deep learning method was developed for high-resolution 2D brain MRF. To adapt for quantitative kidney imaging, multiple network parameters were optimized, including the learning rate policy, patch size and batch size settings. To test whether patient data is needed in the training process, two different training sets were used, including 1) MRF dataset only obtained from normal subjects and 2) dataset obtained from a mixture of normal subjects and patients with pathologies (RCC, ARPKD, and CF). Additionally, to evaluate the performance of deep learning methods in characterizing pathologies, testing dataset was designed to contain data acquired from both normal subjects and patients with different grades of RCC (benign, low-grade and high-grade tumors).

Results

Two different learning rate policies were evaluated, including a ramped learning rate policy and an on-demand policy. For the latter, the learning rate was updated dynamically as the network progressed in the training (Fig 1). Our results show that both the ramped and on-demand learning rates converged to similar minimum loss, but the on-demand policy produced more stable loss curves and converged faster during the training process. We further evaluated network performance with respect to batch and patch sizes. The lowest RMSE values for T₁ and T₂ quantification were achieved with a batch size of 8 and patch size of 64 for 2D kidney MRF (Fig 2).
With the optimized network settings, we further applied the method to retrospectively accelerate kidney MRF with 3x and 4x undersampling, corresponding to tissue mapping using only 576 and 432 time points, respectively. Fig 3 shows that the network trained with patient scans provides slight improvement in comparison to the network trained with normal subjects alone. Fig. 4 shows representative MRF T₁ and T₂ maps obtained from a normal subject using 576 time points. Our results show that deep learning techniques can achieve 3x acceleration along the temporal dimension while providing accurate and high-quality T₁ and T₂ maps of the kidney.
We further evaluated the developed method in quantification of T₁ and T₂ relaxation times for various subtypes of RCCs, including benign, low-grade and high-grade tumors. The evaluation was performed with 3x acceleration and the results were compared with the reference values obtained with full MRF dataset. As shown in Fig 5, the proposed method provides comparable map quality as the reference results using only one third of the MRF data, which corresponds to a short 5-sec MRF scan. Lower T₁ and T₂ values were observed for the low-grade RCC case, which in general presents the highest T₁ and T₂ values. However, no evident differences were noticed for the results with the two training dataset containing patient scans or not.

Discussion and Conclusion

In this study, we evaluated the utility of deep-learning-based tissue mapping for accelerating 2D kidney MRF method. At least 3x acceleration can be achieved, which largely reduces the current 2D kidney MRF scan from 15 sec to only 5 sec, rendering a more comfortable breath-hold that is feasible for most of patients. Our results also suggest that the proposed method can be applied for quantitative T₁ and T₂ characterization of both normal and abnormal tissues in kidney imaging. Future work will be performed to optimize both MRF acquisition pattern and network structure to further improve its quantification accuracy, especially for tissue components with prolonged relaxation times.

Acknowledgements

Our group receives research support from Siemens Healthineers and NIH Grants 1R01CA266702 and R01EB006733.