Scattnet-MR: Interpretable Machine Learning for Accelerated MRI Reconstruction

Scattnet-MR is an interpretable machine learning framework for accelerated MRI reconstruction that combines wavelet scattering transforms with neural implicit representations.

The method is designed to improve robustness and stability under noise and distribution shifts, addressing key limitations of conventional black-box deep learning approaches in medical imaging. By leveraging physically grounded feature extraction, the framework preserves multiscale structure while enabling efficient reconstruction from undersampled data.

This work was carried out at the Laboratorio Avanzado de Procesamiento de Imágenes (LAPI), Facultad de Ingeniería, UNAM.

The manuscript has been submitted to Computers in Biology and Medicine and is currently under review.

Preprint: [Link coming soon]

Summary

• Interpretable ML framework for MRI reconstruction
• Integration of wavelet scattering transforms and neural representations
• Robustness to noise and acquisition variability
• Focus on physically consistent reconstruction

Additional Publications

• Project on cosmological simulation acceleration — coming soon
• Machine learning for physical field reconstruction — coming soon
• Simulation-based inference methods — coming soon

Recommended citation: Villalba, D. (2026). "Scattnet-MR: Interpretable Machine Learning for Accelerated MRI Reconstruction." Computers in Biology and Medicine. Under review.
Download Paper