SSL-QALAS: Self-Supervised Learning for rapid multiparameter estimation in quantitative MRI using 3D-QALAS.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Quantitative Imaging
• Machine Learning in Medical Imaging

Background:

• Accurate multiparametric mapping (T1, T2, proton density, inversion efficiency) is crucial for MRI diagnostics.
• Traditional methods often require extensive data or external dictionaries, limiting speed and accessibility.
• 3D-quantification using an interleaved Look-Locker acquisition sequence with T2 preparation pulse (3D-QALAS) offers comprehensive data but requires efficient processing.

Purpose of the Study:

• To develop and evaluate a self-supervised learning (SSL) based method for rapid, dictionary-free estimation of multiparametric maps from 3D-QALAS measurements.
• To assess the accuracy and generalizability of the proposed SSL-QALAS method compared to traditional dictionary matching.

Main Methods:

• Developed a self-supervised learning based QALAS mapping method (SSL-QALAS).
• Evaluated accuracy by comparing T1 and T2 estimates with reference methods on an ISMRM/NIST phantom.
• Compared SSL-QALAS and dictionary-matching methods in vivo and assessed model generalizability using scan-specific, pre-trained, and transfer learning models.

Main Results:

• Both dictionary-matching and SSL-QALAS methods showed strong linear agreement with reference T1 and T2 values in phantom experiments.
• SSL-QALAS demonstrated comparable performance to dictionary matching for reconstructing T1, T2, proton density, and inversion efficiency maps in vivo.
• Rapid reconstruction (<10s) and fast scan-specific tuning (<15min) were achieved using pre-trained SSL-QALAS models.

Conclusions:

• The SSL-QALAS method enables rapid, dictionary-free reconstruction of multiparametric maps from 3D-QALAS measurements.
• This approach eliminates the need for external dictionaries or labeled ground-truth training data.
• SSL-QALAS offers a promising solution for efficient and accurate quantitative MRI.