In this project, I developed computer vision algorithms and machine learning models for the detection and classification of blood vessels in the anterior surface of the eye.

The goal was to automate the analysis of slit-lamp eye images to support ophthalmologists in diagnostics and research.

Main tasks included:

  • preparing a dataset from medical imagery,
  • segmenting blood vessels on the images,
  • classifying the vessels by their length and thickness.

📈 Key Results and Future Development Link to heading

All the objectives were successfully achieved. The developed algorithms were validated and deployed as a full-featured web service.

For segmentation, a U-Net architecture was used. A pre-trained U-Net model was fine-tuned on our custom dataset, resulting in a segmentation quality of 0.678 (Dice/F1 metric).

Results can be tested on the demo site, where users can upload an eye image and receive segmented and classified vessel visualizations.

The project is planned to evolve further. A portable device is in development, which will allow real-time vessel classification in field conditions.

🏗 Architecture and Development Process Link to heading

The solution architecture includes:

  • Image preprocessing (cleaning, alignment, normalization),
  • Segmentation neural network to extract vessels (UNet architecture),
  • Classifier for determining vessel properties (length, thickness),
  • Web interface for image upload and visualization (built with FastAPI + Vue).

The development process included working with medical data, annotation, model training, and building an accessible frontend for clinical use.

⚙️ Implementation Details, Features, and Challenges Link to heading

Students were involved in annotating the dataset using QuPath. Annotating a single image took around two hours of meticulous work, so the final annotation quality varied significantly. During training, only the highest-quality annotations were used.

We developed post-processing algorithms based on computational geometry to calculate geometric metrics like vessel length, width, and classification. A transformation matrix was applied to improve measurement accuracy, built using a 3D-printed eye model tailored to each microscope used in data acquisition.

Each segmented polygon was decomposed into individual vessels. Then a skeletonized model was constructed, followed by measurement of vessel geometry. Classification was based on empirically derived metrics provided by ophthalmologists.

Thesis Projects Link to heading

The following works were supervised directly by me, where I served as a research advisor: