Akademik Veri Yönetim Sistemi
IEEE Access, 2026 (SCI-Expanded, Scopus)
This study proposes HMA-GPNET, a lightweight and parameter-efficient deep-learning architecture for retinal disease classification based on optical coherence tomography (OCT) images. The network introduces Parallel Fusion Blocks employing decomposed asymmetric convolutions to capture directionally structured retinal features while maintaining low computational complexity. To enhance representational capacity, a dual-stage attention strategy was adopted: Micro-Attention is embedded within the feature extraction hierarchy to refine local features with near-zero additional parameters, whereas Hybrid Multiplicative Attention is applied in the final stage to perform joint channel–spatial recalibration with minimal computational overhead. To assess performance under class imbalance and heterogeneous imaging conditions, the proposed model was evaluated on a unified multi-dataset benchmark integrating six public OCT datasets (OCTDL, DUKE, OCT-C8, OCT2017, NEH, and ARMD-Kaggle) encompassing diverse scanners, acquisition protocols, and patient populations. The experimental results demonstrate competitive classification performance across datasets, although performance varies under domain shift conditions. Despite requiring only 1.546 million trainable parameters and 2.056 GFLOPs, HMA-GPNET achieved dataset-specific accuracies ranging from 92.87% to 100%, depending on dataset characteristics and evaluation conditions. Inference benchmarks on an NVIDIA RTX 3050 Laptop GPU confirmed a GPU latency of 8.18 ms per scan and a model size of 5.91 MB, suggesting potential suitability for resource-constrained environments, subject to further validation on real deployment platforms.