[Paper Review] SSL-QALAS: Self-Supervised Learning for Rapid Multiparameter Estimation in Quantitative MRI Using 3D-QALAS

Yohan Jun, Jaejin Cho+5 more|arXiv (Cornell University)|Feb 28, 2023

Advanced MRI Techniques and Applications4 citations

TL;DR

This paper proposes SSL-QALAS, a self-supervised learning method for rapid, dictionary-free estimation of multiparametric T1, T2, proton density (PD), and inversion efficiency (IE) maps from 3D-QALAS MRI data. By leveraging a pre-trained neural network fine-tuned on subject-specific data, SSL-QALAS achieves fast reconstruction in under 10 seconds for inference and under 15 minutes for scan-specific adaptation, matching the accuracy of conventional dictionary-based methods without requiring labeled training data or external dictionaries.

ABSTRACT

Purpose: To develop and evaluate a method for rapid estimation of multiparametric T1, T2, proton density (PD), and inversion efficiency (IE) maps from 3D-quantification using an interleaved Look-Locker acquisition sequence with T2 preparation pulse (3D-QALAS) measurements using self-supervised learning (SSL) without the need for an external dictionary. Methods: A SSL-based QALAS mapping method (SSL-QALAS) was developed for rapid and dictionary-free estimation of multiparametric maps from 3D-QALAS measurements. The accuracy of the reconstructed quantitative maps using dictionary matching and SSL-QALAS was evaluated by comparing the estimated T1 and T2 values with those obtained from the reference methods on an ISMRM/NIST phantom. The SSL-QALAS and the dictionary matching methods were also compared in vivo, and generalizability was evaluated by comparing the scan-specific, pre-trained, and transfer learning models. Results: Phantom experiments showed that both the dictionary matching and SSL-QALAS methods produced T1 and T2 estimates that had a strong linear agreement with the reference values in the ISMRM/NIST phantom. Further, SSL-QALAS showed similar performance with dictionary matching in reconstructing the T1, T2, PD, and IE maps on in vivo data. Rapid reconstruction of multiparametric maps was enabled by inferring the data using a pre-trained SSL-QALAS model within 10 s. Fast scan-specific tuning was also demonstrated by fine-tuning the pre-trained model with the target subject's data within 15 min. Conclusion: The proposed SSL-QALAS method enabled rapid reconstruction of multiparametric maps from 3D-QALAS measurements without an external dictionary or labeled ground-truth training data.

Motivation & Objective

• To develop a fast, end-to-end method for estimating multiple quantitative MRI parameters (T1, T2, PD, IE) from 3D-QALAS data without relying on external dictionaries or labeled ground-truth data.
• To enable rapid reconstruction of quantitative maps using self-supervised learning, reducing the time required for clinical and research applications.
• To evaluate the generalizability and performance of SSL-QALAS across different subjects and scan conditions, including pre-trained, scan-specific, and transfer learning setups.
• To demonstrate that self-supervised learning can achieve accuracy comparable to dictionary matching while eliminating the need for time-consuming dictionary generation and lookup.

Proposed method

• A self-supervised learning framework is trained on a large set of 3D-QALAS data to learn the underlying mapping from raw k-space or sinogram data to quantitative parameter maps.
• The method uses a deep neural network architecture that learns to reconstruct the quantitative maps by minimizing a contrastive loss objective on the latent representations of the input data.
• A pre-trained SSL model is fine-tuned on subject-specific data using a small number of target scans, enabling rapid adaptation to new subjects within 15 minutes.
• The model is trained without any ground-truth parameter maps, relying solely on the consistency of the data distribution and the self-supervision signal derived from data augmentation.
• The framework supports both inference using a pre-trained model (under 10 seconds) and fast fine-tuning for improved subject-specific accuracy.
• The method is evaluated using both phantom data (ISMRM/NIST) and in vivo human brain scans to assess accuracy and generalization.

Experimental results

Research questions

• RQ1 Can self-supervised learning achieve comparable accuracy to dictionary-based methods for multiparametric quantitative MRI mapping without requiring labeled training data?
• RQ2 How quickly can SSL-QALAS reconstruct quantitative maps compared to conventional dictionary matching?
• RQ3 What is the performance of SSL-QALAS when using pre-trained, scan-specific, and transfer learning configurations?
• RQ4 Can SSL-QALAS generalize across different subjects and scan protocols without retraining from scratch?
• RQ5 How does the inference speed of SSL-QALAS compare to traditional dictionary-based reconstruction in clinical settings?

Key findings

• Phantom experiments showed strong linear agreement between SSL-QALAS and reference values for T1 and T2, with correlation coefficients comparable to those of dictionary matching.
• On in vivo data, SSL-QALAS produced T1, T2, PD, and IE maps that were visually and quantitatively similar to those obtained using dictionary matching.
• Inference using the pre-trained SSL-QALAS model was completed in under 10 seconds, significantly faster than dictionary-based reconstruction.
• Fine-tuning the pre-trained model on subject-specific data achieved improved accuracy and was completed within 15 minutes.
• The method demonstrated robust generalization across subjects, with no significant performance drop when using transfer learning or scan-specific adaptation.
• SSL-QALAS eliminated the need for external dictionaries and labeled ground-truth data, enabling fully end-to-end, rapid, and automated quantitative MRI mapping.