Semi-Supervised Multi-Task Learning for Interpretable Quality As- sessment of Fundus Images

TL;DR

This work addresses the need for interpretable retinal image quality assessment (RIQA) while reducing annotation costs. It proposes a semi-supervised, multi-task framework where a Teacher model trained on a small set provides pseudo-labels for quality details, guiding a Student network to predict both overall RIQA and specific capture conditions. The approach yields improved overall quality classification compared with single-task baselines and produces faithful quality-detail outputs, with GradCAMs that better reflect image defects. The method achieves competitive performance on EyeQ and DeepDRiD and demonstrates that pseudo-label noise can be on par with inter-observer variability, enabling actionable recapture guidance without heavy labeling. The authors also release expert EyeQ detail annotations to support future research.

Abstract

Retinal image quality assessment (RIQA) supports computer-aided diagnosis of eye diseases. However, most tools classify only overall image quality, without indicating acquisition defects to guide recapture. This gap is mainly due to the high cost of detailed annotations. In this paper, we aim to mitigate this limitation by introducing a hybrid semi-supervised learning approach that combines manual labels for overall quality with pseudo-labels of quality details within a multi-task framework. Our objective is to obtain more interpretable RIQA models without requiring extensive manual labeling. Pseudo-labels are generated by a Teacher model trained on a small dataset and then used to fine-tune a pre-trained model in a multi-task setting. Using a ResNet-18 backbone, we show that these weak annotations improve quality assessment over single-task baselines (F1: 0.875 vs. 0.863 on EyeQ, and 0.778 vs. 0.763 on DeepDRiD), matching or surpassing existing methods. The multi-task model achieved performance statistically comparable to the Teacher for most detail prediction tasks (p > 0.05). In a newly annotated EyeQ subset released with this paper, our model performed similarly to experts, suggesting that pseudo-label noise aligns with expert variability. Our main finding is that the proposed semi-supervised approach not only improves overall quality assessment but also provides interpretable feedback on capture conditions (illumination, clarity, contrast). This enhances interpretability at no extra manual labeling cost and offers clinically actionable outputs to guide image recapture.

Figures (7)

• Figure 1: Example use case of the proposed semi-supervised multi-task RIQA, showing the assessment of an initial low-quality image (a) and its improved, recaptured version (b).
• Figure 2: Schematic representation of our proposed semi-supervised multi-task learning approach.
• Figure 3: Per-row normalized confusion matrices for overall quality classification on EyeQ and DeepDRiD test sets, as reported in DenseNet-121-MCF fu2019evaluation, QuickQual engelmann2023quickqual and as obtained by our method.
• Figure 4: Qualitative comparison of results obtained in EyeQ (top) and DeepDRiD (bottom) by the single-task (left) and multi-task (right) models.
• Figure 5: Comparison of our MT-EyeQ model against QuickQual engelmann2023quickqual in image recapture scenarios. Pairs of initial low-quality captures are presented alongside their improved versions from the HRF odstrcilik2013retinal and private datasets.