Deep learning on photoacoustic tomography to remove image distortion due to inaccurate measurement of the scanning radius

Abstract

Photoacoustic tomography (PAT) is a non-invasive, non-ionizing hybrid imaging modality that holds great potential for various biomedical applications and the incorporation with deep learning (DL) methods has experienced notable advancements in recent times. In a typical 2D PAT setup, a single-element ultrasound detector (USD) is used to collect the PA signals by making a 360° full scan of the imaging region. The traditional backprojection (BP) algorithm has been widely used to reconstruct the PAT images from the acquired signals. Accurate determination of the scanning radius (SR) is required for proper image reconstruction. Even a slight deviation from its nominal value can lead to image distortion compromising the quality of the reconstruction. To address this challenge, two approaches have been developed and examined herein. The first framework includes a modified version of dense U-Net (DUNet) architecture. The second procedure involves a DL-based convolutional neural network (CNN) for image classification followed by a DUNet. The first protocol was trained with heterogeneous simulated images generated from three different phantoms to learn the relationship between the reconstructed and the corresponding ground truth (GT) images. In the case of the second scheme, the first stage was trained with the same heterogeneous dataset to classify the image type and the second stage was trained individually with the appropriate images. The performance of these architectures has been tested on both simulated and experimental images. The first method can sustain SR deviation up to approximately 6% for simulated images and 5% for experimental images and can accurately reproduce the GTs. The proposed DL-approach extends the limits further (approximately 7% and 8% for simulated and experimental images, respectively). Our results suggest that classification-based DL method does not need a precise assessment of SR for accurate PAT image formation.

Fig. 1.

Schematic representation of the experimental setup.

Fig. 2.

Demonstration of the dense U-Net (DUNet) architecture.

Fig. 3.

Graphical elaboration of dense block of the first layer of the DUNet.

Fig. 4.

Illustration of ${\mathrm{UNet}}_C$ - image classification by U-Net model.

Fig. 5.

Block diagram detailing image classification and subsequent image formation.

Fig. 6.

Reconstruction results using the simulated images for different approaches for the two-point phantom. Each image, from (b) to (i), represents a normalized reconstructed/predicated image and contains a colorbar quantifying the gray shades. The noise level in the RF signals is higher in the top row than that of the bottom row. The SR is 6% lesser compared to its nominal value (43.75 mm) for the top row and 7% more for the bottom row. The calculated values of the image quality parameters (PSNR and SSIM) are embedded on each image.

Fig. 7.

Performance of various methods on the simulated images for the multi-ellipse phantom.

Fig. 8.

Performance of various methods on the simulated images for the vasculature phantom.

Fig. 9.

Reconstruction results on experimental images for different approaches for the two-point sample.

Fig. 10.

Reconstruction results on experimental images for different approaches for the multi-ellipse sample.

Fig. 11.

Reconstruction results on experimental images for different approaches for the vasculature sample.

Fig. 12.

(a)-(b) Plots of mean absolute error (MAE) for different UNETs during training and validation stages for each epoch. (c) Demonstration of variation of accuracy as a function of epoch. (d)-(f) Same as (a) but for the ${\mathrm{DUNet}}_{TP}$ , ${\mathrm{DUNet}}_{ME}$ , ${\mathrm{DUNet}}_{VA}$ , respectively.