Untrained CNN and coefficient-space regularization for joint reconstruction in photoacoustic tomography

TL;DR

The paper introduces a data-driven method for jointly reconstructing initial pressure and sound speed in photoacoustic tomography using untrained CNN features and coefficient-space regularization. It achieves high accuracy and strong generalization across unseen geometries with significantly reduced computational cost compared to trained CNNs.

Why It Matters

This approach enables faster, more robust image reconstruction in medical imaging without requiring extensive training data or model tuning, making it suitable for real-time clinical applications and diverse anatomical structures.

Key Findings

• Using fixed random convolutional features from an untrained CNN improves reconstruction accuracy over supervised trained-CNNs.
• The method achieves strong generalization to unseen phantom geometries and maintains stability under measurement noise up to 10%.
• Computational time is reduced to about one-sixth of that required by trained CNN approaches while maintaining high reconstruction fidelity.