Noise-aware dynamic convolution for improved generalizability of retinal disease diagnosis using optical coherence tomography images

Abstract

Abstract. Significance Optical coherence tomography (OCT) is widely used for the diagnosis of retinal diseases. However, deep learning models trained on a single dataset often degrade when deployed across scanners and clinical sites due to device-dependent speckle variability and acquisition differences, limiting their reliability in real-world screening. Aim We aim to develop a lightweight deep learning framework that leverages speckle characteristics in OCT images to improve cross-scanner generalizability for retinal disease classification while preserving real-time inference efficiency. Approach We propose NA-DyCNN, a noise-aware dynamic convolutional neural network that minimizes the expected classification risk over multiple stochastic realizations of multiplicative speckle perturbations and regularizes the dynamic routing mechanism to produce scanner-invariant kernel mixtures. The framework was evaluated using over 105,000 B-scans from three heterogeneous OCT cohorts under strict zero-shot cross-dataset transfer. Results NA-DyCNN consistently outperformed lightweight baselines across four zero-shot cross-dataset transfer scenarios, achieving up to 92.87% accuracy, a weighted F2 score of 92.89%, and Cohen’s κ of up to 0.896, demonstrating improved robustness and generalization under cross-dataset shifts. The model maintained high efficiency with only 0.4 M parameters and an inference latency of 0.53 ms per B-scan on an NVIDIA GB10 GPU. Conclusions Modeling postacquisition speckle variability during training improves the generalizability of OCT classifiers without increasing the inference cost, thereby enabling the more reliable deployment of artificial intelligence (AI)-assisted retinal screening across heterogeneous imaging systems.