MULTI-STAGE SEGMENTATION OF RETINAL LAYERS IN OCT B-SCANS

Аннотация

This paper presents a training-free, interpretable, and computationally efficient multi-stage approach for segmenting retinal layers in OCT B-scans. The method is designed to remain robust under speckle noise, depth-dependent illumination nonuniformity, and device/protocol differences, and achieves practical speed on CPU. Clinically meaningful criteria – Dice, IoU, Boundary F1, and layer thickness error – are used for evaluation, alongside pragmatic guidance on parameter selection. The approach delivers reliable and repeatable results while maintaining a “quality–complexity–resource” balance suitable for resource-constrained environments.