Source-linked AI summary
Augmentation Strategy for DINO in Glaucoma Classification
Ashish Kumar Meena, Chandra Sekhar Seelamantula
TL;DR
Natural-image SSL augmentation strategies may not transfer reliably to medical imaging, and DINO’s crop-consistency assumption can be misaligned with fundus anatomy. This paper redesigns DINO’s teacher-student crop assignment for glaucoma-relevant views, reporting stronger performance than alternative assignments and generalization across diverse benchmarks.
Problem
Method
Results
The low-redundancy, task-relevant teacher crops with multi-scale task-retaining student crops consistently outperform alternative teacher-student assignments.
Source: §1 · p. 2Takeaways & Limitations
The learned representations capture anatomically meaningful structure and generalize strongly across diverse benchmarks.
Source: §1 · p. 2The task-relevant crop is currently defined around the optic disc; extending the framework to diseases with more distributed pathology and additional modalities remains future work.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Self-supervised learning (SSL) offers a promising approach to medical image representation learning, yet augmentation strategies of SSL frameworks validated on natural images do not reliably transfer to specialized medical imaging domains. We analyze DINO, a knowledge-distillation based SSL framework, and empirically demonstrate that its default augmentation strategy produces unreliable representations for retinal fundus images, degrading downstream glaucoma classification performance. Guided by the principle that effective cropped views should share minimal redundant information while retaining clinically relevant signal, we redesign DINO's augmentation strategy. Our approach biases the learned representations toward diagnostically critical features rather than low-level image statistics. We evaluate our method for glaucoma classification by training a linear classifier on the frozen representations, achieving strong generalization across seven benchmark datasets. URL for code and weights: https://github.com/ashish-256/task_dino
1 Introduction
Natural-image SSL augmentations may not transfer reliably to medical imaging, and DINO’s crop-consistency objective can suppress glaucoma-relevant fundus structure. The paper proposes task-aware, asymmetric crop assignments and reports stronger performance than alternative assignments.
1 Introduction
- The motivation is that natural-image SSL augmentations do not reliably generalize to specialized medical imaging domains, where optimal augmentation depends on modality and task.Fundus variation across instruments and demographics can lead supervised models to capture dataset-specific artifacts rather than generalizable anatomy.Source: §1 · p. 1
- DINO’s consistency objective can distort glaucoma-relevant structure by enforcing agreement across crops regardless of anatomical relevance.This suppresses features central to glaucoma classification.Source: §1 · pp. 1–2
- The authors redesign DINO’s augmentation strategy with asymmetric teacher-student crop assignment aligned to glaucoma-relevant anatomy.Their contribution addresses a mismatch between DINO’s semantic-consistency assumption and retinal fundus anatomy.Source: §1 · p. 2
- Task-aware, asymmetric teacher-student crop assignments consistently outperform alternative assignments for glaucoma classification.The teacher receives low-redundancy, high task-relevance crops, while the student receives multi-scale crops that retain the task signal.Source: §1 · p. 2
- The learned representations are reported to capture anatomically meaningful structure and generalize across diverse benchmarks.The supplied passage does not specify the benchmark count or metric.Source: §1 · p. 2
2 Related Works
Medical SSL work often transfers natural-image augmentations without adapting them to domain-specific structure. Prior DINO adaptations address task-relevant regions, while this paper examines how teacher and student views should differ.
2 Related Works
- Medical SSL methods have often transferred natural-image augmentation strategies without adapting them to domain-specific structural properties.Examples of domain-tailored approaches include MoCo-CXR for chest X-rays, RETFound for fundus and OCT, and UNI for histopathology.Source: §2 · p. 2
Augmentation Strategies for SSL
- The paper identifies the asymmetric assignment of teacher and student views as an unexamined design question in prior DINO-LG and DINO-EYE work.It systematically evaluates teacher-student crop assignments for fundus imaging.Source: Augmentation Strategies for SSL · p. 2
- DINO’s student learns augmentation-invariant features by matching the teacher’s target distribution across views.Figure 1 depicts this information flow.Source: Figure 1 · p. 3
3 Background: Analysis of DINO Architecture
DINO trains a student to match teacher distributions across image views, relying on crops sharing a consistent semantic identity. Fundus crops can violate that assumption, and the reported similarity analysis finds that default DINO ranks anatomically disjoint crops above connected ones.
3 Background: Analysis of DINO Architecture
- Vanilla DINO trains the student to match teacher probability distributions, with the teacher updated as an exponential moving average of student weights.The teacher processes only global crops, whereas the student processes all views; centering and sharpening help prevent collapse.Source: §3 · p. 3
- DINO assumes all crops of an image share a consistent semantic identity, but fundus crops may contain clinically uninformative, anatomically unrelated regions.This mismatch undermines the reliability of learned representations.Source: §3 · p. 3
- In both REFUGE and ORIGA, RETFound-DINOv2 assigns higher similarity to disc–no-disc crops than to full-image–disc crops, while the task-aware strategy reverses that ordering.RETFound-DINOv2 similarities are 0.58/0.56 versus 0.14/0.09; the task-aware strategy yields 0.57/0.57 versus 0.22/0.35, respectively.Source: §3 · p. 4, Figure 2 · p. 4
4 Methodology: Task-Aware Augmentation Strategy
The method uses an InfoMin-guided asymmetric teacher–student crop assignment to retain glaucoma-relevant information while limiting shared nuisance information. Teacher crops focus on the optic disc; student crops retain it at multiple scales.
4 Methodology: Task-Aware Augmentation Strategy
- The teacher–student view pair minimizes shared information while preserving information relevant to the glaucoma label.The InfoMin framing seeks views that share minimal nuisance information while remaining sufficient for the task.Source: §4 · p. 5
- Teacher and student views use different nuisance allowances: the teacher is designed to retain nearly task-relevant information, while the student retains broader image structure.The teacher’s δ is approximately zero and the student’s δ is much larger; exact constraint satisfaction is described as unattainable in practice.Source: §4 · p. 5
- The teacher receives two vessel-free, minimal-context optic-disc crops, while the student receives three multi-scale crops that each include the disc.All crops use color jitter, flips, and rotations; Gaussian blur is applied only to student crops.Source: §4 · pp. 5–6, Figure 3 · p. 6
5 Experiments and Results
Across seven glaucoma benchmarks, task-aware augmentation yields higher mean AUROC than the two SSL baselines in both linear-probe regimes. Ablations support the asymmetric crop assignment and show a low-data benefit from removing vessels in the teacher crop on PAPILA.
5 Experiments and Results
- Pretraining uses 81k fundus images, and the experiments train a ViT-Small backbone from scratch for 200 epochs with batch size 256.Optic-disc localization and vessel removal are part of the data-preparation pipeline.Source: §5 · p. 6
5.1 Evaluation
- The evaluation uses frozen CLS-token embeddings and linear classifiers across seven benchmarks, with REFUGE as the low-data probe and AIROGS as the large-scale probe.LAG, ORIGA, and FIVES overlap with pretraining data; the authors characterize this as at most transductive bias, not label leakage.Source: §5.1 · pp. 6–8
- Mean AUROC is 0.818 vs. 0.788 for Vanilla DINO and 0.768 for RETFound in the low-data regime, and 0.836 vs. 0.780 and 0.801 in the large-scale regime.The comparison covers seven datasets under both linear-probe regimes; Table 1 reports AUROC with 95% confidence intervals.Source: Table 1 · p. 7, §5.1 · pp. 7–8
5.2 Ablation Studies
- Reverse Views performs worst among the reported variants, with mean AUROC of 0.741 and 0.749, while Common Views scores 0.799 and 0.805.Table 2 summarizes mean ± standard deviation AUROC across the seven datasets for REFUGE- and AIROGS-trained classifiers.Source: Table 2 · p. 8, §5.2 · pp. 8–9
5.1 Evaluation
- Our approach has the best worst-case AUROC across both regimes (0.757 and 0.787), but trails RetiZero on REFUGE in the low-data regime.On CHAKSU, absent from RetiZero’s pretraining distribution, our approach scores 0.784 vs. 0.730 and 0.836 vs. 0.796.Source: §5.1 · pp. 7–8, Table 1 · p. 7
5.2 Ablation Studies
- No Asymmetry scores 0.772 and 0.809; With Vessel’s PAPILA gain from vessel removal is significant in the low-data regime (+0.045, p = 0.002) but not the large-scale regime (+0.029, p = 0.057).The authors interpret this as greater sensitivity of the linear classifier to teacher-side nuisance variation under data-scarce conditions.Source: §5.2 · pp. 8–9
6 Conclusion
The work develops InfoMin-guided, task-aware views for glaucoma classification on color fundus photography. The current task-relevant crop is centered on the optic disc; extension to diseases with more distributed pathology and to other modalities is proposed as future work.
6 Conclusion
- The method is validated for glaucoma classification on color fundus photography, with the task-relevant crop currently defined around the optic disc.The authors identify retinal diseases with more distributed pathology and additional modalities as directions for future extension.Source: §6 · p. 9