APRECOT: New AI method improves optoacoustic imaging with anatomical priors
Deep learning model uses probabilistic anatomy to reconstruct tissue composition with higher accuracy.
A team led by Sarah Franceschin and Dominik Jüstel from the Technical University of Munich has developed APRECOT (Anatomical Priors for Reconstruction of Optoacoustic Tomography), a novel deep learning framework that addresses a fundamental challenge in multispectral optoacoustic imaging: simultaneously segmenting tissues and reconstructing their underlying chemical composition (chromophores). Traditional methods handle these tasks sequentially, causing error accumulation due to the ill-posed nature of the inverse problem where light fluence distributions depend on unknown optical properties. APRECOT overcomes this by conditioning the reconstruction on probabilistic anatomical priors — statistical knowledge about the expected spatial layout and tissue properties derived from existing anatomical databases. This allows the model to infer both tissue boundaries and bulk chromophore concentrations in one integrated pass, dramatically reducing ambiguity.
In proof-of-concept tests using synthetic (in-silico) MSOT data, APRECOT demonstrated marked improvements over reference methods that lack anatomical context or use sequential strategies. The model more accurately estimated chromophore concentrations, particularly for clinically relevant parameters like blood oxygenation levels. This work represents an essential step toward a practical MSOT imaging modality that could directly provide clinically meaningful information — for example, real-time imaging of tissue oxygenation dynamics in tumors or diabetic wounds, or capturing disease-related changes in tissue composition such as fibrosis or inflammation. By embedding deep anatomical priors into the reconstruction pipeline, the researchers have shown that prior knowledge can effectively constrain otherwise intractable inverse problems in biomedical optics.
- APRECOT jointly performs tissue segmentation and chromophore concentration reconstruction in one step, avoiding error accumulation from sequential approaches.
- The method uses probabilistic anatomical priors to guide the ill-posed inverse problem of light fluence distribution in multispectral optoacoustic tomography.
- In silico experiments show APRECOT significantly outperforms reference methods without anatomical context, with potential for real-time monitoring of tissue oxygenation.
Why It Matters
Combines deep learning and anatomical knowledge to enable faster, more accurate optoacoustic imaging for clinical diagnostics like tumor oxygenation mapping.