Exploring uncertainty quantification for photoacoustic image reconstruction and quantitative oxygenation mapping

Deep learning (DL) has recently been employed to enhance photoacoustic (PA) image reconstruction and quantify blood oxygenation. A significant challenge with artificial intelligence (AI) methods is the inability to quantify errors for validating predictions when the ground truth is unknown. Hence, evaluating the predictive reliability of AI models remains a significant obstacle. This study explores uncertainty quantification (UQ) in reconstructing PA images and generating oxygenation maps. 2000 images were simulated with forearm structures at three wavelengths: 750 nm, 800 nm, and 850 nm. We implemented a DNN architecture based on UNet with VGG19 as the encoder and UQ was performed using Monte Carlo Dropout during inference for 10 predictions on both simulated images and real images (in vitro and in vivo images from a volunteer). The input for the DNN architecture was the raw radio frequency (RF) data employing a 128-element linear array, while the targets were the model-based reconstructed image and the simulated oxygenation map. The study indicates that quantitative parameters can be extracted from UQ analysis on DL methods for PA image reconstruction and oxygenation mapping, providing a foundation for improved training strategies and increased robustness in employing DL methods for photoacoustic imaging applications.