MCG-Net: Multiscale Residual Context Learning for Retinal OCT Classification with External Model Comparison

Background and Objectives:Retinal optical coherence tomography (OCT) classification requires recognition of subtle lesions within their anatomical surroundings, yet expanding contextual coverage can obscure local structural information. MCG-Net integrates local morphology and multiscale context within a gated convolutional hierarchy. Methods: A retained depthwise local pathway is augmented by two compressed separable context branches with bounded channel-wise residual scales. This study distinguishes choroidal neovascularization, diabetic macular edema, drusen, and normal retina. OCT-C8 provides development images; independent CAVRI-H5 and OCTDL datasets provide 384 external observations. Both external datasets remain excluded from parameter fitting, development validation, and checkpoint selection, but their ablation results inform selection of the Full 11/23 architecture. Paired comparisons, source-specific assessment, and visual error analysis characterize this external model comparison. Results: MCG-Net achieved 92.01±0.60% accuracy, 91.99±0.61% macro-F1, and 89.47±0.80% Matthews correlation coefficient. Combined context improved macro-F1 over the local baseline by 2.18 percentage points, with a paired 98% interval of 0.50–4.04 and adjusted p=0.019. Exploratory comparisons identified differences against six alternatives, while advantages over the strongest comparators remained uncertain. Source-specific accuracy was 90.86% on CAVRI-H5 and 95.49% on OCTDL; drusen remained the most difficult category. The model contains 26.82 million parameters and has 4.28 ms median inference latency on an RTX 5090. Conclusions: Compressed residual context strengthens gated convolutional OCT classification without replacing local processing. Reuse of the external cohort for architecture selection limits confirmatory interpretation. Untouched external validation, complete-development patient metadata, and representative clinical sampling are needed before clinical use.