Integrating lightweight convolutional neural network with entropy-informed channel attention and adaptive spatial attention for OCT-based retinal disease classification
Md. Rayhan Ahmed, Mohamed Shehata, Patricia Lasserre
Abstract
This article proposes an effective and lightweight contextual convolutional neural network architecture called LOCT-Net for classifying retinal diseases. The LOCT-Net adopts nested residual blocks to capture the local patterns from the optical coherence tomography brightness scans and facilitate gradient flow throughout the network. The multi-scale feature enhancement module incorporates dilation-integrated depthwise strip convolutions to extract fine-grained contextual patterns with an expanded receptive field and a gating mechanism. The extracted features are then refined by a novel feature refinement network consisting of the entropy-informed channel attention module, followed by the adaptive spatial attention module. The entropy-informed channel attention module uses the frequency distribution of pixel values to compute attention weights for spatial analysis. The adaptive spatial attention module focuses on relevant clinical regions and further refines the feature maps in a multi-kernel setting. Additionally, post-explainable artificial intelligence methods are used to provide explanations of LOCT-Net’s decision-making and predictions. The LOCT-Net model has been evaluated on six benchmark datasets, demonstrating an efficient balance between performance and computational cost. With just 2.32 M trainable parameters, the proposed model addresses key challenges in retinal disease classification tasks using OCT B-scans and surpasses previous state-of-the-art methods, achieving an F1 score of 92.98%, 92.34%, 100%, 99.58%, 94.50%, and 97.14% in the OCTID, OCTDL, DUKE, SD-OCT Noor, NEH, and UCSD datasets, respectively. • Context-guided CNN for retinal disease classification with 2.32M parameters. • EICA boosts classification by prioritizing high-entropy channels to capture key features. • Multi-kernel Adaptive Spatial Attention and DsConv capture multi-scale features. • XAI methods like Grad-CAM and LIME provide visual interpretation of decision-making. • Obtains an average accuracy of 96.46% across six benchmark OCT B-scan datasets.