Segmentation of 3D blood vessel networks using unsupervised deep learning

Abstract

Mesoscopic photoacoustic imaging (PAI) enables label-free visualisation of vascular networks in tissue at high contrast and resolution. The segmentation of vascular networks from 3D PAI data and interpretation of their meaning in the context of physiological and pathological processes is a crucial but time consuming and error-prone task. Deep learning holds potential to solve these problems, but current supervised analysis frameworks require human-annotated ground-truth labels. Here, we overcome the need for ground-truth labels by introducing 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 the PAI system in order to learn how to segment vasculature from 3D biomedical images. With a variety of in silico, in vitro and in vivo data, including patient-derived breast cancer xenograft models, we show that VAN-GAN facilitates accurate and unbiased segmentation of 3D vascular networks from PAI data volumes. By leveraging synthetic data to reduce the reliance on manual labelling, VAN-GAN lowers the barrier to entry for high-quality blood vessel segmentation to benefit users in the life sciences applying PAI to studies of vascular structure and function.