SSL Pretraining Improves Selective Prediction in Diabetic Retinopathy Screening
A recent investigation published on arXiv (2605.19133) explores the impact of the duration of self-supervised learning (SSL) pretraining on calibrated confidence and the practice of confidence-based abstention in grading diabetic retinopathy. The study assesses various SSL checkpoints using a consistent fine-tuning approach, focusing on metrics such as calibrated confidence, coverage, selective accuracy, and selective macro-F1. The findings reveal that SSL pretraining enhances selective prediction compared to models trained from the ground up across different datasets and conditions. This research diverges from previous studies that concentrated on downstream accuracy or AUROC, instead examining how the length of SSL pretraining affects confidence behavior in relation to abstention for uncertain predictions, which is critical for screening tasks requiring clinical evaluation.
Key facts
- Study examines SSL pretraining duration effect on calibrated confidence and abstention
- Evaluates multiple SSL checkpoints under fixed fine-tuning protocol
- Metrics include calibrated confidence, coverage, selective accuracy, selective macro-F1
- SSL pretraining improves selective prediction over training from scratch
- Focuses on diabetic retinopathy grading as safety-critical task
- Prior SSL studies mainly evaluate downstream accuracy or AUROC
- Paper available on arXiv with ID 2605.19133
- Research emphasizes need for models to defer uncertain cases to clinicians
Entities
Institutions
- arXiv