Computer-Aided Screening for Eye Diseases from 2-D Retinography

Driss Khalil · June 2022

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

Public retinal-fundus benchmarks (DRIVE, DRIONS-DB, and others)

Report

Thesis PDF

Degree

Master's thesis

University

UniDistance, Brig (Master in AI, jointly with Idiap)

Partnerships

Idiap Research Institute🇨🇭 Switzerland

Several blinding eye diseases develop without any visual symptoms, so a late diagnosis can mean irreversible loss of sight. Screening from two-dimensional retinal photographs offers a cheap way to catch them early, but for a clinician to trust an automated screen, the model should also show its reasoning. That motivated the question behind this thesis: can a single model segment the key fundus structures well enough to support interpretable disease screening?

Automated fundus analysis had largely relied on separate, single-task models, each trained to segment one structure, while the available public datasets were each annotated for only one task. The difficulty, and the hypothesis under test, was that a multi-task model could learn from these disjoint datasets at once by alternating between tasks and filling in the missing annotations with its own predictions from a previous training stage.

A U-Net-based multi-task model was built to segment the two structures most affected by disease, the vessels and the optic disc, and scored with the F1 measure across four experiments varying the balance and the image quality of the disjoint training sets. It underperformed the single-task baselines on images matching the training data, and fell further behind on images of different quality. Two failure modes were isolated and attacked in turn. Training losses were noisy, because alternating between tasks updated the weights many times per epoch; accumulating gradients and updating once per epoch quietened this and improved on the baseline multi-task setup. Convergence, however, remained poor, and filling the missing annotations with the best epoch’s own predictions made it worse still, ending at a higher loss than doing nothing.

The hypothesis was therefore refuted, and instructively so. Because the self-generated labels made matters worse rather than better, poor pseudo-labels cannot be the cause; the diagnostic evidence points instead to the two tasks interfering with one another in a network that shares almost all of its layers between them. The answer to the opening question is thus a qualified no — one model can be trained on disjoint datasets, but not with this degree of parameter sharing — and the thesis names the concrete next step, keeping gradient accumulation while giving each task more of its own layers. The segmentation methods were built on the open-source mednet library, leaving reusable foundations for later fundus-analysis work.