Unsupervised segmentation of 3D microvascular photoacoustic images using deep generative learning

Abstract

Innovations in imaging hardware have led to a revolution in our ability to visualise vascular networks in 3D at high resolution. The segmentation of microvascular networks from these 3D image volumes and interpretation of their meaning in the context of physiological and pathological processes unfortunately remains a time consuming and error-prone task. Deep learning has the potential to solve this problem, but current supervised analysis frameworks require human-annotated ground truth labels. To overcome these limitations, we present an unsupervised image-to-image translation deep learning model called the vessel segmentation generative adversarial network (VAN-GAN). VAN-GAN integrates synthetic blood vessel networks that closely resemble real-life anatomy into its training process and learns to replicate the underlying physics of an imaging system in order to learn how to segment vasculature from 3D biomedical images. To demonstrate the potential of VAN-GAN, the framework was applied to the challenge of segmenting vascular networks from images acquired using mesoscopic photoacoustic imaging (PAI). With a variety of in silico, in vitro and in vivo pathophysiological data, including patient-derived breast cancer xenograft models, we show that VAN-GAN facilitates accurate and unbiased segmentation of 3D vascular networks from PAI volumes. By leveraging synthetic data to reduce the reliance on manual labelling, VAN-GAN lower the barriers to entry for high-quality blood vessel segmentation to benefit imaging studies of vascular structure and function.