UWFound, A domain-adapted foundation model for hierarchical posterior uveitis diagnosis from ultra-widefield fundus images

Technical Summary: UWFound – A Domain-Adapted Foundation Model for Hierarchical Posterior Uveitis Diagnosis

Problem Statement
Accurate diagnosis and aetiological classification of posterior uveitis remain significant clinical challenges due to heterogeneous manifestations, overlapping phenotypes, and severe class imbalance among disease subtypes. Existing artificial intelligence (AI) models face two primary limitations:

1. Modality Mismatch: Most retinal foundation models are pre-trained on conventional colour fundus photographs (CFP), which differ significantly from ultra-widefield (UWF) scanning laser ophthalmoscopy in imaging mechanisms, field of view, colour characteristics, and the appearance of peripheral lesions.
2. Lack of Hierarchical Reasoning: Current models often employ flat classification, failing to reflect the clinical diagnostic workflow where ophthalmologists first detect inflammation, then distinguish between infectious and non-infectious causes, and finally identify specific aetiologies.
   Furthermore, automated UWF-based diagnosis of posterior uveitis is underexplored, and models struggle with rare subtypes due to data scarcity.

Methodology
The authors developed UWFound, a UWF-adapted foundation model integrated into a hierarchical diagnostic framework. The approach consists of three main components:

1. Domain-Adaptive Self-Supervised Fine-Tuning:

   • Base Model: The authors started with RETFound-DINOv2, a generic retinal foundation model.
   • Domain Adaptation: They performed self-supervised fine-tuning using 21,234 unlabelled UWF images collected from routine clinical examinations and public datasets.
   • Preprocessing: Since UWF imaging primarily utilizes red and green laser channels, only these informative channels were retained. Images were resized to 784 × 784 pixels to preserve peripheral retinal and choroidal lesion morphology. This step aimed to bridge the modality gap between CFP and UWF.

2. Hierarchical Diagnostic Framework:
   The model emulates clinical reasoning through a three-level hierarchy:

   • Level 1 (Screening): Distinguishes normal eyes from posterior uveitis.
   • Level 2 (Differentiation): Classifies uveitis as infectious vs. non-infectious.
   • Level 3 (Subtyping): Predicts specific aetiological subtypes (e.g., Behçet's disease, tuberculosis-related uveitis, acute retinal necrosis, various forms of endophthalmitis).
   • Architecture: UWFound serves as the image encoder. A class-conditioned attention module transforms predictions from each level into class embeddings, conditioning subsequent refinement stages to emphasize relevant features while preserving global context. To handle class imbalance and heterogeneity, each level utilizes a prototype-based metric classifier with learnable class prototypes and momentum-updated prototype memory.

3. Training and Evaluation Strategy:

   • Data: The internal labelled dataset comprised 617 UWF images from 373 individuals (265 patients with posterior uveitis, 108 healthy controls). Data was split at the patient level to prevent information leakage.
   • Baselines: Performance was compared against conventional models (ResNet50, ViT) and generic foundation models (VisionFM, RETFound, RETFound-DINOv2).
   • Categories: Evaluation was stratified by data availability: High-support (normal eyes), Medium-support (common subtypes like Behçet's, TB-uveitis, sarcoidosis, ARN), and Few-support (rare subtypes like fungal/bacterial endophthalmitis).
   • External Validation: The model was tested on three independent datasets (WEI-UWF, ZPPH-UWF, and a clinical reference dataset) to assess generalisability.

Key Results

• Screening Performance: In the internal test set, the UWFound-based hierarchical model achieved near-perfect uveitis screening with an AUROC of 0.9999 and an AUPRC of 0.9998.
• Infectious vs. Non-Infectious: The model demonstrated strong discrimination for the second diagnostic level, achieving a mean AUROC of 0.9496 (Non-infectious: 0.9391; Infectious: 0.9098).
• Rare Subtypes (Few-Shot): In medium-support categories, UWFound achieved the highest AUROC (0.8577) and AUPRC (0.4681) among all compared models. In few-support categories, while overall sensitivity remained low (a common challenge for rare diseases), the hierarchical model achieved the highest AUPRC (0.3439), indicating an improved precision–recall trade-off under severe class imbalance.
• External Validation: The model preserved discriminative ability for first-level screening across external datasets (AUROC range: 0.7638–0.9767). However, performance varied significantly based on operating thresholds, highlighting substantial external domain shifts.
• Interpretability: Norm-Layer class activation maps revealed a progressive shift in attention: broad retinal attention at the screening level transitioning to localized lesion-associated activation (e.g., vascular inflammation, peripheral necrotic regions) at finer diagnostic levels.

Significance and Claims
The paper posits that UWFound establishes a robust and generalisable tool for automated posterior uveitis screening and complex aetiological identification. Its primary contributions are:

1. Modality Adaptation: Demonstrating that domain-adaptive self-supervised fine-tuning on large-scale unlabelled UWF data significantly improves representation learning for inflammatory diseases compared to generic retinal foundation models, particularly in medium- and few-shot scenarios.
2. Clinical Alignment: Validating that a hierarchical framework reflecting clinical reasoning (Screening → Infectious/Non-infectious → Aetiology) yields strong performance for the first two critical diagnostic steps, which are essential for treatment decisions (e.g., avoiding corticosteroids in infectious cases).
3. Realistic Utility: The authors modestly claim that while the model provides supportive fine-grained aetiological information, rare-subtype classification remains challenging. Therefore, UWFound is best interpreted as an interpretable triage and decision-support system rather than a standalone diagnostic tool. It is designed to identify eyes likely to have posterior uveitis, prioritize infectious pathways, and suggest potential aetiologies for targeted clinical investigation.

The study concludes that future work must focus on prospective multicentre validation, multimodal integration with clinical and laboratory data, and calibration strategies to ensure safe real-world deployment.