Abstract
Automated glaucoma detection from retinal fundus images plays a crucial role in facilitating early intervention and improving the management of this progressive ocular condition. Although convolutional neural networks (CNNs) have significantly advanced image analysis, current CNN-based models encounter two major limitations. First, they rely primarily on convolutional operations, which restrict the ability to capture cross-channel correlations effectively due to the channel-specific focus of these operations. Second, they often depend on fully-connected (FC) layers for classification, which can introduce unnecessary complexity and limit adaptability, potentially impacting overall classification performance. This study introduces the Modular Convolution-Involution Cascade Network (MCICNet), an innovative CNN architecture designed to address these challenges in the context of glaucoma detection. The model employs a combination of convolution and involution operations in a cascade structure, allowing for the effective capture of inter-channel dependencies within the feature extraction process. Furthermore, the classification phase integrates light gradient boosting machine (LightGBM) as a replacement for traditional FC layers, offering enhanced precision and generalization while reducing model complexity. Extensive experiments conducted on the LAG and ACRIMA datasets demonstrate that MCICNet achieves significant improvements compared to existing CNN and transformer-based models. The model attained a classification accuracy of 95.6% on the LAG dataset and 96.2% on ACRIMA, outperforming nine widely used CNN architectures (AlexNet, MobileNetV2, SqueezeNet, ResNet18, GoogLeNet, DenseNet121, EfficientNetB0, ShuffleNet, and VGG16), as well as three transformer-based models (ViT, MaxViT, and SwinT). Additionally, MCICNet showed superior performance over its variant without involution (MCICNet-NoInvolution). With only 0.9 million parameters, MCICNet demonstrates substantial efficiency in resource utilization alongside its high learning capability, establishing it as an advanced and computationally efficient solution for glaucoma detection.