251 / 2024-08-27 17:32:13

Real-time image registration in fast-scanning photoacoustic microscopy using unsupervised deep learning

Photoacoustic imaging,image registration,deep learning

The inherent scanning distortion caused by fast scanners and dynamic perturbations from high-frequency scans in optical-resolution photoacoustic microscopy systems can lead to significant image misalignments and deformations. These artifacts manifest as misalignment between columns within individual images and deformations between consecutive image sequences, which pose challenges for accurate image analysis. Traditional correction methods typically rely on extracting and matching feature points to rectify these misalignments. However, these methods are often hampered by several limitations, including poor accuracy due to inadequate feature point detection in noisy or low-contrast regions, and low throughput, which restricts their applicability in real-time imaging scenarios. In this paper, we propose an unsupervised deep learning-based approach to restore and register distorted OR-PAM images. Our method leverages a neural network to compute the deformation field between a moving image and a fixed reference image. The computed deformation field is then used to resample the moving image using a differentiable sampler, effectively correcting the distortions and aligning it with the reference image. To evaluate the similarity between photoacoustic images acquired at different time points, we introduce mutual information as a loss function, which is particularly suited for multi-modal and photoacoustic imaging where intensity values do not directly correspond. Our approach was tested on mouse ear data collected using a resonant mirror-based OR-PAM system. The experimental results demonstrate that our proposed method can register up to 40 images per second, which significantly enhances throughput compared to traditional feature-based methods. Moreover, the accuracy of image registration achieved by our method is superior, facilitating more precise extraction of vascular structures and functional information. This is particularly important in biomedical applications. The high speed and accuracy of our approach show its potential for handling the distortions and misalignments inherent in high-speed optical-resolution photoacoustic microscopy imaging, thereby offering a robust solution for advancing fast photoacoustic imaging techniques.