Scale-equivariant deep model-based optoacoustic image reconstruction

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Science

Background:

• Model-based reconstruction is crucial for high-quality multispectral optoacoustic tomography (MSOT) imaging.
• In vivo MSOT data exhibit signal magnitude fluctuations, complicating regularization and supervised deep learning.
• Current methods require manual adjustments for optimal regularization, limiting efficiency and accuracy.

Purpose of the Study:

• To develop a scale-equivariant model-based reconstruction operator for MSOT.
• To enable automatic regularization strength adjustment based on input data.
• To facilitate robust supervised deep learning of reconstruction operators for in vivo MSOT.

Main Methods:

• Derivation of a scale-equivariant model-based reconstruction operator.
• Implementation of automatic regularization adjustment using the L2 norm of the sinogram.
• Training of the deep learning operator with fixed-norm input sinograms.

Main Results:

• The scale-equivariant operator effectively applies regularization to sinograms of varying magnitudes.
• Achieved slightly improved accuracy in quantifying blood oxygen saturation.
• Enabled more accurate supervised deep learning of the reconstruction operator.

Conclusions:

• Scale-equivariant reconstruction offers robust performance for MSOT with fluctuating in vivo data.
• This method enhances the accuracy of quantitative MSOT parameters like blood oxygen saturation.
• It provides a foundation for more reliable deep learning-based MSOT reconstruction.