Contrastive Learning for Accelerated MR Fingerprinting

INTRODUCTION

Magnetic Resonance Fingerprinting (MRF) is a powerful quantitative MRI technique that allows simultaneous estimation of multiple tissue parameters [1]. However, traditional MRF reconstruction methods rely on dictionary matching, which is computationally intensive and noise-sensitive, limiting its scalability for highly undersampled data [2–7]. Although recent methods [8–12] have been proposed to mitigate these issues, practical applications of MRF remain constrained. This study proposes CLIP-MRF, a novel approach leveraging contrastive learning for accelerated, robust MRF reconstruction.

METHODS

Inspired by the Contrastive Language-Image Pre-training (CLIP) [13], we proposed a dual-encoder network that maximizes signal similarity in a learnable space for MRF reconstruction. CLIP is a model for learning associations between images and their corresponding textual descriptions by comparing their embeddings directly. This concept shares common with that of MRF pattern matching. We propose to learn the association between measured signal evolutions and known fingerprint in a dictionary with CLIP. With that, the MRF pattern matching can be replaced by the CLIP network for more accurate quantifications. This is because CLIP compares the features instead of the entire time evolution of measure signals and fingerprints, and thus is more robust than the conventional MRF pattern matching.

The schematic overview of our proposed CLIP-MRF model is shown in Fig. 1(a)(b). The model contains two networks, a signal encoder and a fingerprint encoder respectively. The signal encoder extracts the feature of the measured MRF signal evolutions denoted as, and the fingerprint encoder maps all known fingerprints to a low-dimensional space denoted as. If the features of the two encoders match, then the measured signal represents the corresponding quantitative values of the tissue properties. The blue diagonal denotes the matched pairs whose inner products are maximized, and the inner products of the unmatched pairs are minimized during the training. The right of Fig.1 shows the detailed network architecture.

Figure 1.

Schematic overview of our proposed CLIP-MRF model. The training (a) and the inference process (b) are shown on the left. The network structures of the signal encoder (c) and fingerprint encoder (d) are shown on the right.

Given N input pairs of measured signal evolution and fingerprint of a certain tissue parameter, the signal encoder and fingerprint encoder will be jointly trained to maximize the cosine similarity of the N true pairs in the batch while minimizing that of the rest. This similarity calculation is almost the same as the one in the DM method.

RESULTS & DISCUSSION

The MRF sequence uses a 3D Cartesian trajectory with linear readout in kx and a variable density circular Cartesian undersampling pattern in ky-kz. The MRF acquisition consists of 500 time frames acquired using FISP with each frame acquired at an acceleration factor (AF) of 15. The acquisition matrix size is 96×96×24. Five digital phantoms based on in vivo data from five volunteers under an IRB-approved protocol were used. The in vivo data consist of T1 and M0 maps obtained from 2D multislice phase-sensitive inversion recovery and T2 maps obtained from 3D MAPSS. Quantitative tissue parameter maps were used as input reference values for the generation of MRF data.

To train the CLIP-MRF model, we put the generated MRF data of a pixel into the signal encoder and the fingerprint that corresponds to the T1 and T2 values of this pixel into the fingerprint encoder. The similarity between the features extracted from the two encoders was maximized. Four out of five datasets were used for training and one for testing. The reconstructed maps were compared with conventional dictionary matching (DM) [1], SVD-compressed dictionary matching (LR-DM) [14], and conventional deep-learning-based method (DL) [8]. The same generated MRF data were used to train the DL network. Fig.2 shows the quantification results for a representative sagittal slice. The proposed CLIP-MRF is superior to DM and DL, and comparable to LR-DM but with a significantly less reconstruction time. To demonstrate that image contents have little effect on training, we also trained CLIP-MRF on a synthetic image whose quantification values were randomly assigned for each pixel and tested on the same digital phantom. The results are shown in Fig.3. And Table 1 shows the relative error of all testing data. The same conclusion can be made as in Fig.2, which suggests no need for training datasets with contents similar to the testing ones. We also acquired in vivo MRF data using the same scanner and the same MRF sequence with an AF of 15, and the results are shown in Fig.4. Since there is no ground truth, LR-DM was used as the reference. The CLIP-MRF is the closest to LR-DM in both T1 and T2 maps.

Figure 2.

Comparison of the proposed CLIP-MRF with other networks when trained on the digital phantoms. From left to right: (a) ground truth (b) DM, (c) DL, (d) LR-DM, and (e) proposed CLIP-MRF. A vertical colorbar shows the quantitative values. The nMSE and relative error are shown at the bottom right corner.

Figure 3.

Comparison of the proposed CLIP-MRF with other networks when trained on the synthetic image. From left to right: (a) ground truth (b) DM, (c) DL, (d) LR-DM, and (e) proposed CLIP-MRF. A vertical colorbar shows the quantitative values. The nMSE and relative error are shown at the bottom right corner.

TABLE 1.

The relative errors for T1 and T2 of the reconstruction subject (expressed as mean ± standard deviation, in percentage), defined as the ratio between the Frobenius norm of the difference map and that of the ground truth map.

Method	T1	T2
DM	16.43±5.19	21.92±3.28
LR-DM	16.92±8.43	7.58±2.58
DL	12.29±5.24	15.79±5.84
CLIP-MRF	8.40±4.22	11.98±4.45

Figure 4.

Comparison of the proposed CLIP-MRF with other networks for in vivo MRF data. From left to right: (a) DM, (b) LR-DM, (c) DL, and (d) proposed CLIP-MRF. A vertical colorbar shows the quantitative values. (Top row: T1 maps; Bottom row: T2 maps; units are ms)

CONCLUSION

CLIP-MRF shows improvement over conventional MRF methods. Future work will explore the extension of this framework to other quantitative parameters, such as T2* and B1 inhomogeneity, to further enhance its clinical utility.

Synopsis.

Keywords:

MR Fingerprinting, AI/ML Image Reconstruction, Contrastive Learning

Motivation:

MR Fingerprinting (MRF) enables simultaneous multi-parametric mapping but suffers from computational challenges and noise/artifact sensitivity due to dictionary matching.

Goal(s):

This study aims to develop a novel network called CLIP-MRF to improve pattern matching in MRF. It incorporates contrastive learning to enhance quantification accuracy in accelerated MRF.

Approach:

We propose a dual-encoder contrastive training method to robustly map MRF signals to tissue parameters accurately. The model maximizes similarity between matching signal-parameter pairs and minimizes mismatched ones during training.

Results:

CLIP-MRF demonstrates superior performance over the state-of-the-art MRF methods in T1 and T2 quantification, reducing the reconstruction time and errors.

Impact:

The CLIP-MRF network enables accurate parameter mapping and improves computational efficiency for accelerated MRF. Trained on simulated data only, the network offers robust generalization across signals with different noise/artifacts, paving the way for fast and reliable tissue quantification.