Semantic Segmentation of Weakly Labeled Retinal Images

Xiao Tan · February 2023

Research outputs

mednet. A framework for developing and analysing deep neural network architectures for medical data (2D/3D images and signals), with ready access to many public medical-imaging datasets and extensible model architectures. Software GPL-3.0 · Docs · PyPI · conda-forge · Source

Datasets used

DRIVE, STARE, CHASE-DB1, HRF, IOSTAR (public retinal benchmarks)

Report

Thesis PDF

Degree

Master's thesis

University

University of Zürich, Department of Informatics

Partnerships

Idiap Research Institute🇨🇭 SwitzerlandUniversity of Zürich🇨🇭 Switzerland

Segmenting the vessels in a retinal image means giving every pixel a label, a far more demanding task than classifying the whole image, and one whose progress is capped by how scarce and expensive pixel-level annotations are. This thesis asked whether reliable vessel segmentation can be learned when labelled data is limited, by putting the abundant unlabelled images to work.

Fully supervised networks are bounded by the amount of labelled data, whereas semi-supervised consistency learning promises to exploit unlabelled images as well. The catch is that vessels are thin and sparse, so the data augmentations that consistency methods rely on must be chosen with care. The hypothesis was that extending the Mean Teacher model to retinal images, with augmentations suited to their structure, would let unlabelled data improve segmentation and generalisation.

The Mean Teacher model was adapted to retinal vessel segmentation and the augmentations were searched systematically, with performance measured by the F1 (Dice) score because vessel pixels are so heavily outnumbered by background. Greyscale conversion and additive Gaussian noise proved the effective pair: strong enough to make the two branches disagree usefully, mild enough to leave thin vessels intact, where more aggressive transformations prevented the model from converging at all. Across four public datasets the gains were concentrated exactly where they matter, on the labelled-and-unlabelled pairs that generalise poorly under ordinary supervised training, and were negligible where supervised learning already transferred well. On a genuinely unseen dataset the choice of labelled training set dominated the outcome, and only once that choice was sound did adding further unlabelled data help.

The hypothesis holds, but conditionally. Consistency learning does transfer to retinal vessels and does turn unlabelled images into a real asset, provided the augmentations respect the sparse structure of the target — which answers the opening question affirmatively while naming the constraint that makes it work. The more useful finding is the one that was not anticipated: under weak supervision, choosing the labelled set well matters at least as much as adding unlabelled data, so annotation effort is better spent on which images to label than on how many. The methods live on in the open-source mednet library.