aceNoise-aware dynamic convolution for improved generalizability of retinal disease diagnosis using optical coherence tomography images
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 di…
"CRE-003 OphaGo - Retinal Imaging Adapter - Huced Despro ITS" by Djoko Kuswanto is marked with CC0 1.0. To view the terms, visit https://creativecommons.org/publicdomain/zero/1.0/.
Researchers developed NA-DyCNN, a lightweight noise-aware dynamic convolutional network for OCT-based retinal disease classification that explicitly models post-acquisition speckle variability during training to improve cross-scanner robustness.
By publication on 2026-08-12, the model had been evaluated on over 105,000 B-scans from three heterogeneous cohorts in zero-shot transfer, reaching up to 92.87% accuracy and 0.53 ms latency, suggesting potential for more reliable AI-assisted screening across devices, though real-world clinical deployment and broader scanner diversity remain to be validated.
- Proposed NA-DyCNN minimizes expected classification risk over stochastic multiplicative speckle perturbations and regularizes dynamic routing for scanner-invariant kernel mixtures.
- Evaluation used over 105,000 B-scans from three heterogeneous OCT cohorts under strict zero-shot cross-dataset transfer across four scenarios.
- 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.
Deep learning OCT models trained on a single dataset often degrade across scanners and sites due to device-dependent speckle variability, limiting reliability in real-world screening.
The rundown
OCT is widely used for retinal disease diagnosis, but models trained on one dataset degrade across devices due to speckle variability and acquisition differences.
Authors developed NA-DyCNN, a lightweight noise-aware dynamic CNN that models multiplicative speckle perturbations during training and regularizes dynamic routing to produce scanner-invariant kernel mixtures.
Tested on over 105,000 B-scans from three heterogeneous cohorts in four zero-shot transfer scenarios, NA-DyCNN outperformed lightweight baselines and maintained 0.4M parameters with 0.53 ms per B-scan latency on an NVIDIA GB10 GPU.
Sources
- Peer-reviewedJournal of Biomedical Optics2026-08-12
How should this claim be treated?
The debate