HSI

Automated Retinal Vessel Segmentation in OCT Videos

We built a machine learning model that automatically finds blood vessels in high-resolution OCT scans and, where possible, distinguishes arteries from veins. This reduces manual work for clinicians and helps spot circulation issues in the eye faster.

Promotional project image

Initial Situation

Eye diseases like retinal vessel occlusion or glaucoma can lead to vision loss if not detected early. A new Swiss method makes blood flow visible in time-resolved OCT videos. Today, experts still mark vessel centers manually—time-consuming and inconsistent. Our aim: automate reliable vessel detection and support artery/vein labeling.

Problem statement / Project goal

We designed a simple, practical AI pipeline that:

  • segments vessels in OCT volumes automatically,
  • explores artery vs. vein classification using image features (intensity profiles and temporal signals),
  • evaluates denoising to improve signal quality while keeping vessels sharp.
gallery-1
image description
image description
image description

Solution developed and its benefits

We built an end-to-end pipeline that turns raw OCT videos into clear vessel maps and helpful artery/vein hints. It combines reliable segmentation with simple image cues and careful denoising to reduce manual effort while keeping results understandable.

  • Automatic vessel finding: a self-configuring segmentation model detects vessel regions and centers consistently across scans.
  • Human-readable cues: intensity and reflectivity profiles (and first pulse signals) provide intuitive hints for artery vs. vein.
  • Denoising: Non-Local Means and a lightweight temporal window average reduce speckle noise but preserve vessel walls.
  • Time savings and consistency: fewer manual clicks, more uniform outcomes—useful for triage, study evaluation, and follow-ups.

In practice, the system already detects arteries reliably and reduces manual work. Veins are harder (lower contrast, fewer labels), but the denoising and emerging temporal features lay a solid foundation for further gains.

Key terms

  • OCTOptical Coherence Tomography - high-resolution cross-section images of the retina.
  • SegmentationLabel each pixel/voxel to delineate vessels or tissues.
  • nnU-NetSelf-configuring deep learning framework for medical image segmentation.
  • DenoisingNoise reduction (e.g., Non-Local Means, Wavelet, temporal window averaging).
  • DiceOverlap score 2*TP/(2*TP+FP+FN); punishes thin structures.
  • Intensity profileBrightness across a vessel — arteries tend to be brighter than veins.

Customer

Augenklinik Basel Logo

University of Basel - Eye Clinic
Project partners: PD Dr. med. Peter Maloca, PhD Dr. med. Philippe Valmaggia
Mittlere Str. 91
4056 Basel
www.augenklinik.usb.ch

Team

Denis Wilhelm, Livio Brunner

Advisor
Prof. Dr. Susanne Suter
Simone Bonato