Synopsis

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction, Magnetic Resonance Cholangiopancreatography

MR cholangiopancreatography (MRCP) is a special MRI technique to visualize the biliary systems. Deep Learning-based (DL) reconstruction models have shown to reduce scan time from many anatomical regions. However, they generally require large training datasets. This is challenging for applications like MRCP, where public datasets are not available. This work analyzes two approaches to training a DL model for highly accelerated MRCP reconstruction: training from scratch using a small MRCP dataset and fine-tuning a model pretrained on a public knee dataset. Results show that despite of the substantial data domain shift between training and testing, fine-tuning outperformed training from scratch.

Introduction

MRCP is an MRI technique that visualizes the biliary systems using a heavily T2-weighted sequence. Acquiring MRCP is challenging due to irregular breathing, causing undiagnostic images. Therefore, reducing scan times is very important for the MRCP exam, and techniques like parallel imaging^1-2 (PI) and compressed sensing³ (CS) have already been employed successfully for MRCP⁴. DL methods have shown promising results in MRI reconstruction at high accelerations⁵. Unfortunately, they generally require large training datasets, challenging for applications like MRCP due to the absence of public datasets. Additionally, the characteristics of MRCP are substantially different from public datasets, such as a fastMRI⁶ dataset, and there are significant anatomical variances⁷ between subjects. It is an open research question if a pretraining with a public dataset with such a substantial domain shift can improve the performance of DL reconstruction models. This abstract analyzes two training strategies for DL reconstruction models for highly accelerated MRCP: 1) a model pre-trained on a public dataset and fine-tuned with MRCP data 2) a model trained from scratch using only MRCP data. The results from both DLs were also compared against L2-sensitivity encoding² (SENSE) and L1-Wavelet CS qualitatively and quantitatively.

Method

Data acquisition and retrospective undersampling
MRCP data were acquired using a 3-D T2-weighted turbo spin-echo⁸ (SPACE) sequence from 14 volunteers on a 3 T MR scanner (MAGNETOM Vida, Siemens Healthcare, Erlangen, Germany) using a clinical routine protocol with coherent undersampling and an acceleration rate of three (PAT3) with 24 autocalibration signal (ACS). Furthermore, prospective acquisition correction technique⁹ (PACE) triggering reduced motion artifacts. 3-D MRCP k-space was inverse Fourier-transformed along the slice direction to enable independent 2-D slice reconstructions. The 3-D volume was normalized in a range of absolute values of zero to one by dividing by a maximum absolute value (MAV). We reconstructed the PAT3 data with GRAPPA¹ to serve as the ground truth for our DL trainings. We then performed PAT9 undersampling retrospectively with 24 ACS.

DL model architecture and trainings
We used the variational network⁵ (VN) with identical architecture of 12 cascades and approximately 32M parameters for all DL trainings. In contrast to the standard VN data processing pipeline⁵, instead of using a zero-filled image as the input, we performed an initial SENSE reconstruction for the network input.
Pretraining was performed using a subset of 77/15 randomly selected training/validation examples from the fastMRI multi-coil knee dataset (Fig. 1a). Fine-tuning was then performed using 10/2 training/validation examples from our MRCP dataset with a reduced learning rate (Fig. 1b). Training from scratch was performed using the same 10/2 MRCP training/validation datasets with the same learning rate as during pre-training stage (Fig. 1c).
Two MRCP examples were left out as the test set to evaluate the performance of each model. Table 1 contains detailed information about each model, and Figure 1 shows an overview of the pipeline.

Reference CS and SENSE
We used the Sigpy python package¹⁰ to generate reference L2-SENSE and L1-Wavelet CS. We used ESPIRiT¹¹ for receive coil sensitivity estimation with 24 ACS and two thresholds for k-space and image, θ_k = 0.02 and θ_image = 0.9, respectively. The regularization parameters λ for L2-SENSE and L1-Wavelet CS were set such that the metric scores to the ground truth were maximized for the MRCP test set. This resulted in the values λ_SENSE = 0.02 and λ_CS = MAV. MAV depends on the MRCP datasets.


Evaluation
MRCP results are usually displayed as maximum intensity projection (MIP) images to facilitate the assessment of the entire biliary system. For quantitative assessment, we calculated the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) to the ground truth images. Each slice was normalized between zero and one for the quantitative evaluation. Both metrics were calculated slice by slice and averaged across slices afterwards. We defined the visibility of small image details as our main criterium for qualitative evaluation.

Results

The fine-tuned model (PSNR [dB]: 34.72/SSIM [%]: 94.68) and the MRCP model trained from scratch (PSNR: 34.01/SSIM: 94.09) lead to the best results on the test set (Figs. 3 and 4). Both models contain less noise and aliasing artifacts compared to the L2-SENSE (PSNR: 29.47/SSIM: 81.79) and L1-Wavelet CS (PSNR: 31.64/SSIM: 86.23) reconstructions. The high degree of undersampling also leads to wavelet compression artifacts in L1-Wavelet CS (Figs. 3 and 4, white arrows) and to residual aliasing artifacts in L2-SENSE which can be mistaken as ducts (Figs. 3 and 4, white arrowheads). Pre-training without fine-tuning leads to the poorest performance (PSNR: 22.69/SSIM: 52.53) of all reconstructions.

Discussion and Conclusion

In this work, we evaluated the performance of the DL reconstruction model for highly accelerated MRCP. Our results show that in the case of substantial domain shifts between train and test data, the performance of DL models depends crucially on the availability of at least a small data set from the target application. Using only public datasets without fine-tuning for model training leads to worse results than standard CS and parallel imaging. However, it is remarkable that even when training a DL model from scratch with such a small dataset, DL outperforms CS and parallel imaging in terms of quantitative image metrics and qualitative inspection of the images.

Acknowledgements

No acknowledgement found.