Answers are grounded in this paper and cite their sources.
The optimal neural network architecture for classifying retinal diseases from optical coherence tomography (OCT) images remains unclear. We evaluated a proposed attention-enhanced deep-learning framework, CDBA-RetNet, and compared multiple backbone architectures and attention mechanisms for OCT image classification. Six backbone architectures and seven attention mechanisms were evaluated using five-fold patient-level cross-validation on the OCTDL dataset, which contains 2064 B-scan images from 821 patients across six retinal diseases and normal controls. Cross-dimensional bottleneck attention (CDBA) generates separate height, width, and channel attention maps through bottleneck transformations and combines them using multiplicative fusion. Among the attention configurations evaluated with the ResNeXt101 backbone under the same protocol, CDBA-RetNet achieved the highest mean test macro F1-score of 88.8% ± 3.3% and accuracy of 93.7% ± 1.6%. Relative to the ResNet50 reference, test macro F1-score increased by 2.8%, with the largest class-level F1-score gains observed for epiretinal membrane (5.9%), diabetic macular edema (5.5%), and vitreomacular interface disease (5.2%). In selected examples, Grad-CAM qualitatively highlighted anatomically plausible retinal regions. These findings support further investigation of CDBA-RetNet for retinal OCT classification.
No takes yet. Share an insight, caveat, or question.
Dokumaci et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: