Improving microvascular brain analysis with adversarial learning for OCT-TPM vascular domain translation

Area of Science:

• Biomedical Engineering
• Neuroscience
• Medical Imaging

Background:

• Microscopic cerebrovascular network modeling is crucial for understanding cerebral blood flow and oxygen transport.
• Optical Coherence Tomography (OCT) and Two-Photon Microscopy (TPM) are key imaging modalities for microvascular analysis.
• TPM offers superior localization and image quality but is limited by dye leakage issues, unlike OCT.

Purpose of the Study:

• To develop a method for generating high-quality TPM-like angiographies from OCT vascular data using generative adversarial learning.
• To compare the effectiveness of 2D and 3D CycleGAN models in reconstructing vascular structures from OCT.
• To evaluate the generated vascular structures for image quality and topological accuracy.

Main Methods:

• Utilized generative adversarial learning, specifically 2D and 3D CycleGANs, trained on unpaired OCT and TPM vascular image datasets.
• Evaluated generated TPM angiographies based on image similarity and signal-to-noise ratio.
• Assessed the 3D topological accuracy of generated vascular networks after segmentation and model extraction.

Main Results:

• The 2D CycleGAN model achieved superior image quality compared to the 3D model.
• The 3D CycleGAN model demonstrated consistent superiority in accurately generating vascular network features and topology.
• The generative approach successfully produced high-quality TPM angiographies from OCT data.

Conclusions:

• Generative adversarial learning offers a viable method to enhance vascular analysis using OCT data, especially when TPM is impractical.
• The 3D CycleGAN model is more effective for preserving the intricate topological features of cerebrovascular networks.
• This technique provides a valuable complementary approach for detailed vascular analysis in neuroscience and biomedical research.