Investigation of a learned image reconstruction method for three-dimensional quantitative photoacoustic tomography of the breast

Abstract

Quantitative photoacoustic computed tomography (qPACT) holds great promise to advance a variety of important clinical applications with its potential to estimate vital physiological properties such as oxygen saturation. However, the qPACT reconstruction problem is highly nonlinear and ill-posed. Conventional spectral unmixing methods often oversimplify the problem, resulting in suboptimal accuracy. Alternatively, more principled image reconstruction approaches that comprehensively model the imaging physics are computationally burdensome and require the design of effective regularization strategies. To overcome these limitations, learning-based methods have been proposed. To date, however, the effectiveness of such methods on full-scale problems in which clinically relevant variability in anatomy and physiological parameters is considered has not been established. To address this, we investigated the use of a convolutional neural network with spatial and channel attention modules to estimate the three-dimensional (3D) distribution of tissue oxygenation within vessels and lesions in the female breast. The network was provided with input data comprising noise-corrupted 3D initial pressure distributions corresponding to three wavelengths (757, 800, 850 nm). An additional novel aspect of our study was the use of realistic 3D numerical breast phantoms that described stochastic variations in breast anatomy and functional properties, which enabled a meaningful, quantitative, and systematic evaluation of the proposed method. This study represents an important contribution to the field of qPACT and will guide the exploration of learning-based methods to help translate this important technology by delineating potential prospects and limitations.