OCT-VisionNet: A CBAM Integrated NasNet Architecture for Retinal Disease Classification From OCT Images With Diverse Dataset Validation

Accurate and reliable prediction of retinal diseases from Optical Coherence Tomography (OCT) images is vital for early diagnosis and clinical decision making. Most of the CNN models achieved better classification performance, but failed to capture the small lesion variation from the OCT images. Implementing the traditional deep learning models in a clinical setting becomes uncertain due to its high model size, parameters and inference time. The proposed OCT-VisionNet model bridges this gap by integrating data augmentation and Convolutional Block Attention Module (CBAM) with the NasNetMobile architecture. After preprocessing and data augmentation, the NasNetMobile model extracts features and generates feature maps from OCT images. The CBAM is then placed only between the feature extractor and the global average pooling layer, which specifies the network’s ‘what’ and ‘where’ to focus on the feature map, thus highlighting lesion-focused classification. The proposed model was evaluated using three datasets. The accuracy achieved for Dataset 1, Dataset 2 and Dataset 3 was 80%±4.7%, 95%±0.5% and 97%±0.0% and the small lesion localization was visualized using GradCAM, an explainable AI. The classification accuracy for all datasets increased over the baseline NasNetMobile model by 36%, 10% and 29%, respectively, which was statistically ensured using McNemar’s statistical analysis. The model size, parameters and inference time of the proposed model were 51.5 MB, 5.98 M and 0.3 sec, which were comparatively low when compared with traditional deep learning models. This framework improves the prediction accuracy and also ensures robustness and transparency, which are important requirements for clinical deployment.