Supervisor: Professor Huaxiong Huang, Department of Mathematics and Statistics, York University
Co-Supervisors: Professor Shixin Xu, Department of Mathematics, Duke Kunshan University; Faisal Habib, MADS Lab
Project Location: Fields Institute (Toronto, ON)
Project Overview
Retinal images provide a non-invasive way to assess the health of the retina as well as infer possible health issues in other parts of the body. Vasculature health is one of the key indicators, which can be quantified by measuring properties such as vessel width, length, curvature, and connectivity. These measurements often begin with an automated segmentation algorithm that identifies the vessels in an image. However, even small segmentation errors can produce unexpectedly large changes in the measurements calculated from the resulting vessel network. This project will use existing retinal vessel segmentation methods, previously generated segmentation results, and publicly available retinal image datasets to study how such errors propagate into quantitative vascular features. The student will introduce controlled changes to vessel masks, measure how different vascular features respond, and investigate mathematical ways to describe and reduce this sensitivity while preserving meaningful vessel structure. The expected outcomes are a systematic understanding of which vascular measurements are most sensitive to segmentation errors, improved guidelines for obtaining stable measurements, reproducible computational tools, and a final research report and presentation.
---
Retinal vessel segmentation is widely used as an intermediate step in quantitative fundus-image analysis. Once a vessel mask has been obtained, it can be transformed into centerlines, vascular graphs, local vessel diameters, curvature measurements, and other quantitative representations. These quantities can subsequently be used to characterize vascular morphology or as inputs to downstream statistical and classification models.
Our recent study evaluated several existing retinal vessel segmentation pipelines on multiple public retinal image datasets. The analysis considered pixel-level agreement, thin-vessel recovery, structural fidelity, and downstream vascular features. One important finding was that segmentation quality measured at the pixel level is not sufficient for ensuring the reliability of quantitative vascular features computed using the extracted vessel masks.
In particular, a binary vessel mask can be transformed into a skeleton, a vascular graph, and a local diameter representation. Small changes in the original mask can remove thin branches, create or break graph connections, alter local diameter estimates, and strongly affect curvature or tortuosity measurements. Controlled perturbation experiments further demonstrated that different vascular features can respond very differently to the same small changes in a segmentation mask.
These observations motivate the central research question of this project: How do segmentation errors propagate through different mathematical representations of a vascular network, and how can vascular features be constructed so that they remain stable under small segmentation perturbations while preserving meaningful vessel structure?
The project will address this question using existing segmentation algorithms, existing segmentation outputs, and existing public retinal-image datasets. It is not the aim of this project to develop new retinal vessel segmentation algorithms. Instead, we will focus on the mathematical and computational analysis of the following pipeline: Segmentation Mask → Vascular Representation → Quantitative Feature.
The main objective is to understand and quantify the stability of this pipeline and to identify feature processing procedures that reduce artificial sensitivity to segmentation errors without removing meaningful vascular information.
A second research question is whether sensitive vascular features can be made more stable without destroying meaningful biological structure?
Several possible feature-processing strategies will be investigated using existing mathematical and computational tools. These may include spline-based centerline representations, averaging or regularization of continuous diameter estimates, and removal of short graph spurs. The purpose of these operations is not to make the vascular network artificially smooth, but rather to reduce numerical amplification of small segmentation errors.
No new clinical or experimental data collection is required. The project will use existing publicly available retinal-image datasets and previously generated segmentation results. The current benchmark includes the following public datasets: AFIO, CHASEDB1, DualModal2019, HRF, LES, STARE.
Existing vessel annotations and segmentation outputs from previously evaluated retinal vessel segmentation pipelines will serve as the starting point for the project. The student will therefore not be responsible for developing or training a new segmentation model. Instead, existing methods and their predictions will be treated as inputs to the mathematical analysis. This allows the majority of the summer program to be devoted to sensitivity analysis, mathematical formulation, computational experiments, and interpretation of results.
Expected Outcomes
By the conclusion of the nine-week program, students are expected to produce:
- a quantitative framework for measuring how segmentation errors propagate into vascular features;
- a systematic sensitivity analysis covering several classes of segmentation perturbations and vascular representations;
- an evaluation of feature-processing strategies designed to improve stability while preserving meaningful vascular structure;
- cross-dataset evidence identifying which observations are robust across different retinal-image datasets and existing segmentation methods;
- reproducible computational code and visualizations; and
- a final research report and presentation summarizing the mathematical formulation, computational findings, and remaining open questions.
More broadly, the project may provide a general framework for studying the reliability of quantitative measurements derived from imperfect image segmentation. Although retinal vascular imaging provides the motivating application, the underlying question of error propagation and feature stability arises in many areas of computational imaging.
Student Responsibilities
The student will be responsible for:
- working with existing retinal-image datasets and previously generated segmentation results;
- implementing controlled perturbations of vessel masks;
- computing vascular representations including masks, centerlines, graphs, and diameter maps;
- quantifying the sensitivity of vascular features to different segmentation errors;
- investigating the trade-off between feature stability and preservation of vascular structure;
- conducting numerical experiments across multiple datasets and existing segmentation methods;
- preparing visualizations and analyzing representative success and failure cases;
- maintaining reproducible computational code; and
- preparing the final research report and presentation.
The supervisor will provide guidance on the mathematical formulation, experimental design, interpretation of vascular measurements, and presentation of the research results.
Key Tasks and Nine Week Timeline
- Week 1: Introduction to retinal vascular representations; become familiar with the existing datasets, segmentation outputs, and feature-extraction pipeline. Reproduce selected existing results.
- Week 2: Implement controlled perturbations of existing vessel masks and establish a consistent computational framework for measuring perturbation magnitude and feature changes.
- Week 3: Study sensitivity of area-, length-, and diameter-based vascular measurements under different perturbations.
- Week 4: Study error propagation through skeletonization and graph representations, with particular attention to vessel connectivity, branch points, and terminal branches.
- Week 5: Study sensitivity of geometric features such as curvature and tortuosity. Compare the effects of different perturbation types and perturbation magnitudes.
- Week 6: Develop and compare quantitative measures of feature sensitivity. Identify which vascular representations amplify segmentation errors and investigate the mechanisms responsible for the amplification.
- Week 7: Investigate stable feature representations using existing processing techniques such as centerline smoothing, diameter averaging, and graph pruning. Analyze the trade-off between robustness and preservation of vascular structure.
- Week 8: Perform cross-dataset and cross-segmentation-method validation to determine which conclusions are consistent across different retinal-image datasets and existing segmentation outputs.
- Week 9: Analyze representative success and failure cases; finalize reproducible computational code, figures, the final research report, and the research presentation.