Joint segmentation and image reconstruction with error prediction in photoacoustic imaging using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning enhances photoacoustic (PA) image reconstruction but struggles with error quantification for validation, especially with limited-bandwidth ultrasonic detectors.
• Accurate validation is crucial for quantitative PA imaging applications.

Purpose of the Study:

• To propose a novel Hybrid Bayesian Convolutional Neural Network (Hybrid-BCNN) for joint PA image and segmentation prediction with uncertainty quantification.
• To enable robust validation of PA image reconstruction even when ground truth is unknown.

Main Methods:

• Developed a Hybrid-BCNN model integrating Bayesian principles with convolutional neural networks.
• Trained the model using simulated PA data, incorporating segmentation to focus on signal-rich regions.
• Applied the trained model to both simulated and experimental PA data.

Main Results:

• The Hybrid-BCNN accurately predicts PA images and segmentations.
• Quantified error predictions demonstrated a high statistical correlation with actual reconstruction errors.
• Confidence-based processing using error predictions improved the quality of reconstructed PA images.

Conclusions:

• The Hybrid-BCNN effectively addresses the challenge of error quantification in PA image reconstruction.
• This approach enhances the reliability and quantitative accuracy of PA imaging, particularly in data-limited scenarios.
• Uncertainty prediction is a valuable tool for validating and improving deep learning-based PA imaging.