A robust lightweight deep learning network with deformable convolutions for efficient segmentation of optical...

Manual segmentation and evaluation of optical coherence tomography (OCT) images are difficult and rely on specialist training, which limits the output in a busy clinic. While automatic segmentation can ease this burden, the deep models that currently lead the field tend to be too heavy for routine deployment on standard clinical hardware. A further obstacle is the unusual geometry of OCT data: instead of compact objects on a background, the structures of interest form long, undulating tissue strata that extend horizontally across the whole frame, and this layout is poorly matched to convolutional kernels designed for ordinary photographs.

We address both issues with LiDeOCTNet, a compact segmentation backbone tailored to the layered nature of OCT. Within a reduced UNet, we introduce learnable sampling offsets so that each kernel can reshape itself locally and trace the curved, irregular interfaces between tissues. To assess the design, we evaluate the network on three OCT collections covering retinal as well as endoscopic acquisitions.

Across all three, LiDeOCTNet surpasses a parameter-matched UNet baseline and approaches the accuracy of much heavier benchmarks such as TransUNet, despite using about 750× fewer parameters and running roughly 7.4× faster at inference.

Together these properties point to LiDeOCTNet as a practical option for real-time OCT analysis pipelines, including embedded and resource-limited devices.