Zero-Shot Artifact2Artifact: Self-incentive artifact removal for photoacoustic imaging without any data

TL;DR

This work tackles reconstruction artifacts in 3D photoacoustic imaging caused by sparse, angle-limited detectors by introducing Zero-Shot Artifact2Artifact (ZS-A2A), a zero-shot self-supervised framework that learns artifact patterns through randomized perturbations of the input data. A two-subset perturbation strategy paired with a lightweight CNN $g_{\theta}$ predicts and subtracts artifacts by minimizing a symmetric residual loss and a consistency loss, with final artifact-free reconstruction given by ${clean} = {recon} - g_{\tilde{\theta}}({recon})$. Across simulations and in vivo rat studies, ZS-A2A achieves state-of-the-art performance among zero-shot methods, significantly improves CNR (e.g., rat liver ${17.48} \rightarrow {43.46}$ and kidney ${21.56} \rightarrow {46.31}$), and runs efficiently (${\approx}$8 s per slice and ${\approx}$25 min for full 3D on high-end GPUs). The approach eliminates the need for external training data and prior artifact knowledge, offering a practical and scalable solution for improving 3D PAI quality and potentially extending to other imaging modalities.

Abstract

Photoacoustic imaging (PAI) uniquely combines optical contrast with the penetration depth of ultrasound, making it critical for clinical applications. However, the quality of 3D PAI is often degraded due to reconstruction artifacts caused by the sparse and angle-limited configuration of detector arrays. Existing iterative or deep learning-based methods are either time-consuming or require large training datasets, significantly limiting their practical application. Here, we propose Zero-Shot Artifact2Artifact (ZS-A2A), a zero-shot self-supervised artifact removal method based on a super-lightweight network, which leverages the fact that reconstruction artifacts are sensitive to irregularities caused by data loss. By introducing random perturbations to the acquired PA data, it spontaneously generates subset data, which in turn stimulates the network to learn the artifact patterns in the reconstruction results, thus enabling zero-shot artifact removal. This approach requires neither training data nor prior knowledge of the artifacts, and is capable of artifact removal for 3D PAI. For maximum amplitude projection (MAP) images or slice images in 3D PAI acquired with arbitrarily sparse or angle-limited detector arrays, ZS-A2A employs a self-incentive strategy to complete artifact removal and improves the Contrast-to-Noise Ratio (CNR). We validated ZS-A2A in both simulation study and $ in\ vivo $ animal experiments. Results demonstrate that ZS-A2A achieves state-of-the-art (SOTA) performance compared to existing zero-shot methods, and for the $ in\ vivo $ rat liver, ZS-A2A improves CNR from 17.48 to 43.46 in just 8 seconds. The project for ZS-A2A will be available in the following GitHub repository: https://github.com/JaegerCQ/ZS-A2A.

Figures (6)

• Figure 1: Randomized perturbations-based artifact patterns learning. (a) Applying randomized perturbations to reconstruction results by discarding data from randomly selected detectors. (b) Artifact removal based on the learned artifact patterns from randomized perturbations.
• Figure 2: Numerical instability of artifact fluctuations under random perturbations. (a) Subset sensor data generated from 200 independent random discarding perturbations. (b) Coefficient of variation (CV) of artifacts and PA signals in the reconstructed image.
• Figure 3: The overview of Zero-Shot Artifact2Artifact pipeline.
• Figure 4: Comparison of results using different zero-shot artifact removal methods. (a) Artifact removal results for image of complex vessel-MAP. (b) Artifact removal results for images of complex vessel-slice. (c) Artifact removal results for image of simple phantom-MAP. (d) Artifact removal results for images of simple phantom-slice. (Scale: 2.0 mm.)
• Figure 5: 3D PA reconstruction results of a rat liver. (a) XY Plane-MAP, XZ Plane-MAP, YZ Plane-MAP and the cross-section slice at green dashed line and red dashed line marked in XZ Plane-MAP of the UBP 3D reconstruction results using 1,024 sensor signals. (b) XY Plane-MAP, XZ Plane-MAP, YZ Plane-MAP and the cross-section slice at green dashed line and red dashed line marked in XZ Plane-MAP of the UBP 3D reconstruction results after ZS-A2A artifact removal using 1,024 sensor signals. (Scale: 2 mm.) (c) Schematic diagram of the imaging area. (d) 3D volume display of the reconstruction results using maximum intensity projection.