BiOrthoNet : Efficient Retinal OCT Classification Using Direction‐Specific Convolutions

ABSTRACT Reliable interpretation of optical coherence tomography (OCT) scans is crucial for the early detection of retinal diseases. Deep learning has emerged as a promising approach for this task; however, many existing models are computationally intensive and fail to exploit the horizontal‐vertical structural organization commonly observed in retinal OCT images. This work introduces BiOrthoNet, a lightweight convolutional architecture that incorporates directional feature extraction inspired by retinal anatomy via strictly parallel anisotropic convolutions and an efficient BiOrthoAttention module that adaptively fuses orthogonal feature streams. Despite its compact size of 2.24 million parameters and a computational cost of 1.93 GFLOPs, BiOrthoNet achieves consistent performance across multiple OCT benchmarks with varying data distributions and acquisition conditions, reporting accuracies of 99.79%, 98.00%, 97.82%, 94.16%, 92.75%, and 92.24% on OCT2017, UCSD, OCT‐C8, NEH, OCTDL, and OCTID, respectively. The OCT2017 accuracy is reported under the dataset's official train/test partition for benchmark comparability, while a supplementary leakage‐free evaluation under a strict global patient‐level re‐split (five random seeds) yielded accuracy, which is considered a more reliable estimate of generalization. The proposed model also delivers low‐latency inference speeds of 5–14 ms per B‐scan, making it suitable for efficient inference under limited computational resources. Furthermore, the generated directional saliency maps highlight horizontal and vertical retinal structures, indicating that the model captures directionally consistent structural patterns in OCT images. All experiments were conducted on a workstation equipped with an NVIDIA RTX 3050 GPU (6 GB VRAM) and an AMD Ryzen 7 5235HS CPU. Overall, BiOrthoNet offers an efficient and anatomically motivated solution for automated OCT‐based retinal disease classification.