Application of a Virtual Imaging Framework for Investigating a Deep Learning-Based Reconstruction Method for 3D Quantitative Photoacoustic Computed Tomography

TL;DR

This work tackles the challenge of validating a deep-learning–based reconstruction for 3D qPACT by introducing a realistic virtual-imaging (VI) framework built on stochastic 3D breast phantoms and high-fidelity forward models. A representative DL qPACT method is trained to jointly estimate $sO_2$ maps and vascular/tumor segmentation from multispectral input, and is evaluated across idealized (Study 1) and realistic (Study 2) imaging conditions. Results show strong $sO_2$ estimation and localization performance under ID conditions, with robust generalization to unseen skin tones in Study 1, but notable degradation in segmentation and tumor-detection performance under realistic acoustic variability and OOD skin tones in Study 2. The findings underscore the value of VI-based validation for uncovering method limitations, guiding data-diversity strategies, and informing subsequent in vivo validation in the development of robust qPACT reconstruction methods.

Abstract

Quantitative photoacoustic computed tomography (qPACT) is a promising imaging modality for estimating physiological parameters such as blood oxygen saturation. However, developing robust qPACT reconstruction methods remains challenging due to computational demands, modeling difficulties, and experimental uncertainties. Learning-based methods have been proposed to address these issues but remain largely unvalidated. Virtual imaging (VI) studies are essential for validating such methods early in development, before proceeding to less-controlled phantom or in vivo studies. Effective VI studies must employ ensembles of stochastically generated numerical phantoms that accurately reflect relevant anatomy and physiology. Yet, most prior VI studies for qPACT relied on overly simplified phantoms. In this work, a realistic VI testbed is employed for the first time to assess a representative 3D learning-based qPACT reconstruction method for breast imaging. The method is evaluated across subject variability and physical factors such as measurement noise and acoustic aberrations, offering insights into its strengths and limitations.

Figures (13)

• Figure 1: Distributions of functional, acoustic, and optical properties of a representative type B NBP with an embedded malignant tumor: (a) blood oxygen saturation $sO_2$, (b) speed of sound $c$, (c) optical scattering coefficient $\mu_s$ at a wavelength of 757 nm, (d) optical absorption coefficient $\mu_a$ at 757 nm, and (e) 3D malignant tumor model. For visualization purposes, the tumor is shown as a split volume in (e). The inset in (d) displays a cross-section with arrows indicating the tumor locations. Volumetric renderings were generated using ParaView ParaView, and color maps were manually adjusted to enhance visual clarity. These anatomically realistic numerical phantoms provide a versatile and clinically meaningful platform for developing and evaluating qPACT techniques under realistic physiological and anatomical variability.
• Figure 2: Virtual imaging system configuration. (a) Arc-shaped light delivery subsystem composed of linear fiber-optic segments; (b) schematic of a custom line beam with conical angular emission from a single fiber-optic segment; (c) all effective ultrasonic transducer positions from the rotating arc‐shaped array around the breast across 480 tomographic view steps louisapark2023stochastic.
• Figure 3: Overview of the dual-task DL network for simultaneous $\text{sO}_2$ estimation and anatomical segmentation in 3D photoacoustic tomographic images. Three reconstructed initial pressure distributions at illumination wavelengths of 757, 800, and 850 nm serve as inputs to a shared encoder. Two separate decoders then generate (i) a whole-breast $\text{sO}_2$ map and (ii) a binary segmentation map restricted to the outermost 1.5 cm shell from the breast surface, where veins, arteries, and tumors (if present) regions are labeled as 1 and all other voxels as 0. A combined loss function compares the predicted outputs with the corresponding ground truth maps ($\text{sO}_2$ maps and segmentation masks).
• Figure 4: Visual comparison of estimated blood oxygen saturation ($\text{sO}_2$) distributions and segmentation maps of vessels and tumors for the ID test set (skin color 1) in Study 1. Top row: estimated $\text{sO}_2$ maps obtained using spectral unmixing, fluence-compensated unmixing, and DL-based qPACT (left to right), each masked using the corresponding estimated segmentation map. Bottom row: true segmentation map (left), estimated segmentation map (center), and true $\text{sO}_2$ masked with the estimated segmentation map (right). DL-based qPACT provided more consistent $\text{sO}_2$ maps and more accurate segmentation.
• Figure 5: Visual comparison of DL-based qPACT results for the OOD test sets (skin colors 3 and 5) in Study 1. Top row: estimated (first and third) and true (second and fourth) $\text{sO}_2$ maps, each masked with the corresponding estimated segmentation map, for skin color 3 (first and second) and skin color 5 (third and fourth). Bottom row: estimated segmentation masks for skin colors 3 (left) and 5 (center), and the corresponding true segmentation mask (right). DL-based qPACT maintained high visual fidelity in both segmentation and blood oxygenation estimates across diverse skin tones, demonstrating robust generalization.