Myelin segmentation in cross-circular polarization birefringence microscopy
OA Version
Citation
Abstract
Quantifying myelin in three dimensions is critical for studying neurodevelopment, pathology, and repair. However, dense manual annotation of high-resolution microscopy z-stacks is often impractical. This thesis develops a scalable segmentation pipeline for Cross-Circular Polarization Birefringence Microscopy (CCP-BRM) brain volumes that reduces reliance on extensive manual labels.We generate weak supervision using two interpretable heuristics: rolling-ball background subtraction and Tenengrad (squared-gradient) focus measure. These capture complementary aspects of the myelin signal to produce pseudo-label masks for in-focus fibers. We then train a context-aware 2.5D U-Net that leverages stacked neighboring slices to predict a center-slice mask. Increasing axial context from K=1 to K=5 yields modest improvements in probabilistic segmentation quality (AUPRC) on a development crop.
We further fine-tune the selected model using a small set of densely annotated slices to better match the manual target definition. When evaluated on a physically distinct volume, fine-tuning improves AUPRC relative to both the pretrained model and the heuristic baseline. Pixel-wise overlap metrics remain moderate, reflecting the inherent challenge of segmenting sparse, thin structures in this modality. Finally, we demonstrate the downstream utility of these masks for conservative myelin volume density estimation and qualitative 3D visualization. Overall, this work establishes a practical workflow for CCP-BRM myelin fiber segmentation and highlights the data requirements for stronger generalization in future 3D myelin analysis.
Description
2026