Invariant representations in deep learning for optoacoustic imaging.

Area of Science:

• Medical Imaging
• Computational Science
• Machine Learning

Background:

• Optoacoustic tomography (OAT) image reconstruction is sensitive to measurement parameters.
• Existing algorithms may fail when applied to settings different from their training configuration.
• Developing robust and invariant reconstruction methods is crucial for practical OAT applications.

Purpose of the Study:

• To explore deep learning for invariant and robust OAT image reconstruction.
• To investigate the efficacy of the ANDMask scheme for OAT.
• To assess the out-of-distribution generalization of deep learning models in OAT.

Main Methods:

• Application of deep learning algorithms focused on invariant and robust representations.
• Utilizing the ANDMask scheme adapted for the OAT inverse problem.
• Conducting numerical experiments with variations in parameters like sensor locations.

Main Results:

• Deep learning models trained for invariance and robustness show no performance degradation with out-of-distribution data.
• In some cases, explicit invariance/robustness training even improved performance compared to standard deep learning.
• The ANDMask scheme proved adaptable and effective for OAT reconstruction.

Conclusions:

• Deep learning approaches incorporating invariance and robustness are highly effective for OAT image reconstruction.
• These methods enhance generalization capabilities, making OAT more reliable across different experimental setups.
• The findings support the development of more adaptable and universally applicable OAT reconstruction algorithms.