Towards robust quantitative photoacoustic tomography via learned iterative methods
New AI approach cuts training data needs while improving image reconstruction accuracy...
Researchers Anssi Manninen, Janek Gröhl, Felix Lucka, and Andreas Hauptmann have published a new paper on arXiv proposing a learned iterative approach for quantitative photoacoustic tomography (QPAT). Traditional PAT reconstruction methods rely on computationally expensive forward models and sufficient prior information to handle noisy measurements, making them slow for time-critical applications like real-time surgical guidance. While deep learning alternatives offer faster reconstructions, they typically require large training datasets that are impractical in medical imaging.
The team's method integrates model-based iterative updates—specifically gradient descent, Gauss-Newton, and Quasi-Newton methods—directly into neural network training. They explored both greedy training (optimizing each iteration separately) and end-to-end training (jointly optimizing all networks). Testing on ideal simulated data and a digital twin dataset that mimics real-world scarcity and modeling errors, the approach demonstrated superior generalizability with limited training data, significantly outperforming standard deep learning models. This work bridges the gap between classical physics-based reconstruction and modern AI, potentially accelerating PAT's clinical adoption for high-resolution tissue imaging without the need for massive annotated datasets.
- Learned iterative methods combine physics-based models with neural networks for QPAT reconstruction
- Gradient descent, Gauss-Newton, and Quasi-Newton updates were tested with both greedy and end-to-end training
- Digital twin dataset validated robustness against scarce training data and high modeling errors
Why It Matters
Enables faster, more accurate photoacoustic imaging with limited data, advancing real-time clinical diagnostics.