Uncertainty-Aware Last-Layer Adaptation of RETFound for Referable Diabetic Retinopathy Screening Under Dataset Shift

Self-supervised learning (SSL) is now a standard way to pretrain medical image models, but performance is still mostly judged by downstream accuracy. For safety-critical screening tasks such as diabetic retinopathy grading, this is not enough: a model must also know when its predictions are unreliable and defer uncertain cases for clinical review. In this work, we examine how the length of SSL pretraining influences calibrated confidence and confidence-based abstention. We evaluate multiple SSL checkpoints under a fixed fine-tuning protocol and assess calibrated confidence, coverage, selective accuracy, and selective macro-F1. Across datasets and data regimes, SSL pretraining improves selective prediction compared to training from scratch. Unlike prior SSL studies that primarily evaluate downstream accuracy or AUROC, we analyze how SSL pretraining duration influences confidence behavior under calibrated confidence-based abstention. However, once accuracy saturates, selective performance can still change markedly across checkpoints, and longer pretraining does not consistently improve reliability. These results underscore the importance of abstention-aware evaluation and suggest that pretraining length should be treated as an important reliability-related design choice rather than only a computational detail. Code is available at GitHub.

The quality of diabetic retinopathy (DR) screening relies on the ability to correctly grade severity; however, many deep-learning (DL) classifiers cannot be easily interpreted in the clinical context. This study presents a methodology that combines strong discriminative models with multimodal explanations, converting retinal pixels into clinically interpretable outputs. Using the APTOS 2019 benchmark, we evaluated six representative CNN- and transformer-based backbones under a controlled protocol with stratified five-fold cross-validation. We then compared ensembling strategies (hard voting, weighted soft voting, stacking) and investigated a hybrid class-level fusion variant to exploit grade-specific advantages. For interpretability, we produced Grad-CAM++ visual attribution maps and short textual rationales using vision-language models (VLMs) conditioned on the fundus image and classifier outputs under conservative prompting constraints. Modern CNN backbones (ResNet-50 and ConvNeXt-Tiny) provided the strongest single-model baselines, with cross-validated QWK up to 0.919 and 0.914, respectively. Ensembling improved ordinal agreement, and weighted soft voting was the most consistent across folds (QWK 0.934 +/- 0.017). Hybrid class-level fusion was competitive but did not yield a statistically reliable improvement over standard fusion in paired fold comparisons (Holm-adjusted p >= 1.000). For explanation quality, Grad-CAM++ offered plausible but coarse localization, and VLM rationales were generally grade-consistent. Quantitatively, VLM variants showed a trade-off between clinical completeness and template-level semantic similarity (coverage 0.700 vs. BERTScore 0.072), while image-text alignment was comparable (CLIPScore approximately 0.34).

Purpose: Early screening for eye diseases is critical in low- and middle-income countries where access to care is limited. We investigate whether a confidence-guided, multi-image diabetic retinopathy diagnosis framework can integrate image filtering with confidence-aware predictions for reliable screening at capture. Methods: We develop a multi-image fusion method that aggregates retinal views to improve confidence and balanced accuracy. Our method uses confidence to identify unreliable predictions, prompting retakes when needed. We compare: (1) a cascaded image-quality and disease diagnosis pipeline using a single image per patient, (2) confidence-based prediction, and (3) our confidence-based multi-image fusion pipeline. All methods are evaluated using a RETFoundGreen backbone on the mBRSET (n = 1,234) and BRSET (n = 7,599) datasets. Results: At 70% coverage, our method achieves 91% balanced accuracy on mBRSET and 97% on BRSET, improvements of ~12% and ~6%, respectively, over cascade filtering. The image-quality cascade reaches sensitivities of 61% on mBRSET and 86% on BRSET, whereas our framework reaches 94% and 96%, respectively, at 50% coverage. Conclusions: Human-annotated quality labels are weakly associated with diagnostic performance, and confidence-based filtering consistently outperforms image quality-based cascaded pipelines. Translational Relevance: Using confidence-based multi-image fusion, patients receive more reliable predictions, reducing incorrect diagnoses during screening. The lightweight backbone and single inference pass per image make the framework compatible with low-latency mobile screening systems in resource-limited settings.