Hybrid CNN-Transformer for Retinal OCT hits 95.4% accuracy with 12x better calibration

Animesh Kumar's new paper introduces a Calibrated Hybrid CNN-Transformer for retinal OCT classification, tackling a critical gap in medical AI: models that are accurate but can't communicate when they're wrong. The encoder fuses convolutional and Transformer layers to capture both local texture and global structure in OCT scans, then feeds into a gradient-boosting (XGBoost) classification head. On a four-class dataset of 84,495 retinal OCT scans, the model reaches 95.4% accuracy—competitive with state-of-the-art—but its real innovation is trustworthiness: a three-part clinical safety layer that includes confidence calibration (cutting expected calibration error twelve-fold to 0.0024), out-of-distribution (OOD) rejection for unknown pathologies, and per-prediction uncertainty flags.

This is the first OCT classifier to validate all three safety mechanisms together, making it safer for clinical deployment. False-positive or false-negative predictions in retinal imaging can delay sight-saving treatment, so knowing when to trust a model—and when to escalate to a human—is paramount. Kumar has released the full code, model weights, and a REST inference API on GitHub and HuggingFace, supporting reproducible multi-seed evaluation. The work underscores a growing trend: accuracy alone isn't enough for medical AI; calibrated uncertainty and OOD detection are essential for real-world adoption.