Deep learning-based quantitative optoacoustic tomography of deep tissues in the absence of labeled experimental data

Deep learning (DL) shows promise for quantitating anatomical features and functional parameters of tissues in quantitative optoacoustic tomography (QOAT), but its application to deep tissue is hindered by a lack of ground truth data. We propose DL-based "QOAT-Net,"which functions without labeled experimental data: A dual-path convolutional network estimates absorption coefficients after training with data-label pairs generated via unsupervised "simulation-to-experiment"data translation. In simulations, phantoms, and ex vivo and in vivo tissues, QOAT-Net affords quantitative absorption images with high spatial resolution. This approach makes DL-based QOAT and other imaging applications feasible in the absence of ground truth data.