Neurovascular Segmentation in sOCT with Deep Learning and Synthetic Training Data

TL;DR

This work tackles the challenge of segmenting cerebral microvasculature in high-resolution, label-free sOCT images by proposing a synthesis-based training pipeline. It introduces a spline-based synthetic label engine and a texture-rich synthetic image generator to train a 3D U-Net entirely on synthetic data, leveraging high-variance domain randomization to generalize across acquisitions. The approach achieves human-level segmentation accuracy, demonstrates robustness via extensive ablations and baselines, and scales to full cortical volumes in practical time on standard GPUs. The results suggest synthetic data can overcome labeling bottlenecks in vascular imaging and enable high-throughput vascular network mapping with potential applications in neurovascular disease research.

Abstract

Microvascular anatomy is known to be involved in various neurological disorders. However, understanding these disorders is hindered by the lack of imaging modalities capable of capturing the comprehensive three-dimensional vascular network structure at microscopic resolution. With a lateral resolution of $<=$20 {\textmu}m and ability to reconstruct large tissue blocks up to tens of cubic centimeters, serial-section optical coherence tomography (sOCT) is well suited for this task. This method uses intrinsic optical properties to visualize the vessels and therefore does not possess a specific contrast, which complicates the extraction of accurate vascular models. The performance of traditional vessel segmentation methods is heavily degraded in the presence of substantial noise and imaging artifacts and is sensitive to domain shifts, while convolutional neural networks (CNNs) require extensive labeled data and are also sensitive the precise intensity characteristics of the data that they are trained on. Building on the emerging field of synthesis-based training, this study demonstrates a synthesis engine for neurovascular segmentation in sOCT images. Characterized by minimal priors and high variance sampling, our highly generalizable method tested on five distinct sOCT acquisitions eliminates the need for manual annotations while attaining human-level precision. Our approach comprises two phases: label synthesis and label-to-image transformation. We demonstrate the efficacy of the former by comparing it to several more realistic sets of training labels, and the latter by an ablation study of synthetic noise and artifact models.

Figures (9)

• Figure 1: Visualization of sOCT data and manual segmentation of the vasculature. Left panel: A slice overview of an sOCT acquisition with regions of interest (a, b, c) outlined in red, each showing different anatomical features. Scale bar (1cm) shown in white in bottom left corner. Right panels: detailed manual segmentation for the corresponding regions viewed in axial, sagittal, and coronal planes, overlaid with segmented vessels (green). Scale bar (500 µm) shown in black in bottom right corner.
• Figure 2: Overview of proposed method. Labels and intensity textures are synthesized to resemble volumetric sOCT data which is used to train a U-Net in binary vessel segmentation. We then use these models for large-scale prediction and compare with expert labelers.
• Figure 3: Non-exhaustive demonstration of texture mapping procedure for three separate instances of combined label maps.
• Figure 5: Qualitative comparison of model performance in different regions of normal control samples taken from the human somatosensory cortex. Subfigure (a) illustrates 3D renderings of TPs (green) and FPs (yellow) from the prediction of our proposed model (A). Subfigure (b) shows the same information but as 2d sections and with the addition of FNs (magenta). Scale bar (1000 µm) shown in black in upper left corner.
• Figure 6: DSC, FPR, and FNR for synthesis conditions A-H (n=3) on human somatosensory cortex with respect to expert human labeler. Mean values shown at top of bar, error bars represent standard deviation.