Compressed Sensing MRI with ℓ₁-Wavelet Reconstruction Revisited Using Modern Data Science Tools.

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging (MRI)
• Artificial Intelligence

Background:

• Deep learning (DL) excels in accelerated MRI reconstruction, often outperforming traditional methods like compressed sensing (CS).
• Conventional CS methods typically rely on a few empirically tuned hyperparameters, unlike the extensive parameterization in DL.
• There is a need to enhance traditional methods with advanced data science techniques for improved performance and interpretability.

Purpose of the Study:

• To revisit and optimize ℓ1-wavelet CS for accelerated MRI using modern data science tools.
• To demonstrate that optimized CS can achieve performance comparable to DL methods.
• To develop a more explainable reconstruction model with fewer parameters.

Main Methods:

• Utilized algorithm unrolling and end-to-end training with stochastic gradient descent, mirroring DL techniques.
• Integrated conventional concepts such as wavelet sub-band processing and reweighted ℓ1 minimization.
• Trained the model over large databases, leveraging data science principles.

Main Results:

• Achieved accelerated MRI reconstruction quality comparable to DL methods using an optimized ℓ1-wavelet CS approach.
• The proposed method employs significantly fewer parameters (128) compared to DL (hundreds of thousands).
• The reconstruction model is fully explainable and convex.

Conclusions:

• Modern data science tools can significantly enhance traditional CS methods for accelerated MRI.
• Optimized ℓ1-wavelet CS offers a competitive and more interpretable alternative to DL in MRI reconstruction.
• This approach provides a powerful, explainable, and parameter-efficient solution for accelerated MRI.