Neural networks for faster laser ultrasound tomography in tissue phantoms

Speed of sound (SoS) mapping provides quantitative and localised information about a material's acoustic properties, allowing identification of spatial compositional changes. In biomedical applications, SoS variations can inform tissue characterisation or improve image reconstruction algorithms that typically assume a constant SoS. However, conventional time-of-flight (ToF) tomography methods remain computationally intensive. This study presents experimentally derived tomographic reconstructions of SoS maps of heterogeneous structures from all-optically acquired data using a convolutional neural network (CNN). The CNN, trained on simulated data, enables near real-time, high-quality tomographic reconstructions. The novelty of this work lies in the integration of a laser ultrasound (LU) data acquisition setup with a CNN-based reconstruction approach, demonstrating its potential for remote and flexible inspection of biomedically relevant samples. The CNN was trained using simulated data based on ultrasonic wave propagation models and achieved tomographic reconstructions of a 77 mm $\times$ 77 mm area in less than 6 ms. Data were acquired from four tissue-mimicking phantoms (30 mm diameter) with inclusions of varying size (minimum 6 mm diameter) and SoS (minimum variation 25 m/s). When compared with published, in vivo studies using mammography (MM), conventional ultrasound, and magnetic resonance imaging (MRI), the proposed method yielded 5.73% mean sizing error for phantoms and inclusions relative to the ground truth, highlighting the clinical potential of the LU-CNN framework and the need for further in vivo testing. These findings underscore the method's potential as a precise, faster alternative to conventional imaging and reconstruction methods used in clinical practice.