RADU-Net: Eliminating the Necessity of Measurement of Precise Scanning Radius for Image Reconstruction in Photoacoustic Tomography

The use of deep convolutional neural networks (DCNNs) in photoacoustic tomography (PAT) has grown over the past few years. A circular scanning geometry is typically used for 2D PAT involving a single-element transducer for signal capture. The exact scanning radius (SR) is vital for efficient image reconstruction. Here, we propose a DCNN-based radius-adjusted dense U-Net(RADU-Net) to accurately reconstruct images without relying on the exact SR as input The inputs to DCNN were 128x128 pixel images, and the output images were of the same size. It was trained with 800 heterogeneous images of two distinct phantoms generated using the backprojection (BP) method. The forward data for those phantoms were generated using the k-Wave toolbox. The effectiveness of the RADU-Net was tested using phantom experiments employing a Q-switched Nd:YAG laser (532 nm wavelength and 7 mj/cm ² fluence) and a single-element transducer (2.25 MHz center frequency, 70% bandwidth) having SR=29.80 mm. For a ±5% error in the original SR, we observed that the BP images are completely distorted, but a faithful recovery of the ground truth (SSIM≈1) is possible with the RADU-Net.