TezNet: An Efficient Convolutional Neural Network for Optical Coherence Tomography Classification

Retinal diseases such as Choroidal Neovascularization (CNV), Drusen, and Diabetic Macular Edema (DME) can severely impair vision and are challenging to reverse if detected late. Optical Coherence Tomography (OCT) serves as a non-invasive imaging modality to diagnose these conditions through cross-sectional retinal images. Current advancements in retinal disease detection predominantly focus on fine-tuning larger models to achieve high accuracy, often at the expense of substantial training time and data requirements. We introduce an innovative Convolutional Neural Network (CNN) named TezNet that is deeper and wider than existing custom CNNs for the dataset used in this research yet maintains a lightweight architecture with only 3.6 million parameters. This high efficiency is accomplished by leveraging depthwise and spatial separable convolutional techniques without employing preprocessing techniques or noise reduction filters. The data augmentation techniques generate diverse image qualities, ensuring robust predictions across various scenarios. By training exclusively on a single dataset, our model achieves rapid training times and optimal resource utilization, minimizing the risk of overfitting. This blend of advanced parameter optimization, minimal preprocessing, and strategic data augmentation underscores the potential of our CNN model for swift, precise, and resource-efficient retinal disease detection. TezNet model achieves a remarkable 96% accuracy, competing closely with state-of-the-art networks like ResNet50 (over 23M parameters), InceptionV3 (over 25M parameters), and ResNeXt (89M parameters).