Unsupervised Segmentation of 3D Microvascular Photoacoustic Images Using Deep Generative Learning

Area of Science:

• Biomedical Imaging
• Medical Image Analysis
• Deep Learning

Background:

• Mesoscopic photoacoustic imaging (PAI) provides high-contrast, label-free visualization of vascular networks.
• Accurate segmentation of 3D vascular networks from PAI data is critical for understanding tissue physiology and pathology.
• Current segmentation methods are often time-consuming, error-prone, and require extensive manual annotation.

Purpose of the Study:

• To develop an unsupervised deep learning framework for automated 3D vascular network segmentation from PAI data.
• To reduce the reliance on manually annotated ground-truth labels in PAI analysis.
• To create a model capable of segmenting vasculature by learning the physics of PAI.

Main Methods:

• Introduction of the Vessel Segmentation Generative Adversarial Network (VAN-GAN), an unsupervised image-to-image translation model.
• Integration of synthetic blood vessel networks resembling real anatomy into the training process.
• Training VAN-GAN to replicate the underlying physics of the PAI system for segmentation.

Main Results:

• VAN-GAN demonstrated accurate and unbiased segmentation of 3D vascular networks across diverse datasets (in silico, in vitro, in vivo).
• The model was successfully applied to patient-derived breast cancer xenograft models and 3D clinical angiograms.
• Achieved competitive segmentation performance (F1 score: 0.84) compared to supervised methods like U-Net (F1 score: 0.87).

Conclusions:

• VAN-GAN offers a robust, unsupervised solution for segmenting 3D vascular networks from PAI data.
• The use of synthetic data and physics-informed learning lowers the barrier for high-quality vascular segmentation.
• This approach has the potential to significantly enhance preclinical and clinical research on vascular structure and function.