Retinal disease classification from spectral domain optical coherence tomography using lesion attentive fusion

Retinal diseases such as central serous chorioretinopathy (CSC) and neovascular age-related macular degeneration (nAMD) can cause permanent vision loss, and their diagnosis from three-dimensional spectral-domain optical coherence tomography (SD-OCT) volumes is challenging because lesions may appear in only a few critical slices. We developed a lesion-attentive fusion (LAF) model that automatically weights pathological slices for volume-level retinal disease classification. The model, comprising a slice-wise encoder, LAF layer, and classifier, was evaluated on 641 SD-OCT volumes (16,025 B-scans) from 589 patients with nAMD or CSC subtypes or normal retinas. In eight-fold cross-validation, it achieved 84.85% accuracy and an F1-score of 82.16%, outperforming baseline fusion models. On a selected test fold, the model also outperformed all six independent ophthalmologists. Its slice-level attention agreed with expert lesion-slice annotations (Cohen’s kappa = 0.67, p < 0.001), and Grad-CAM highlighted disease-relevant regions. We additionally evaluated cross-disease applicability using 626 UK Biobank OCT volumes from 462 patients across five retinal disease categories with patient-level five-fold cross-validation. ConvNeXt-Tiny with LAF achieved the highest macro F1-score of 0.7500 ± 0.0583. These results suggest that lesion-attentive aggregation enables accurate and interpretable classification of retinal diseases from OCT volumes.