Towards efficient glaucoma screening with modular convolution-involution cascade architecture.

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Automated glaucoma detection from retinal fundus images is vital for early intervention.
• Current Convolutional Neural Network (CNN) models struggle with cross-channel correlations and complex classification layers.
• Limitations include restricted inter-channel dependency capture and potential performance impacts from fully-connected layers.

Purpose of the Study:

• To introduce the Modular Convolution-Involution Cascade Network (MCICNet) for improved glaucoma detection.
• To address limitations in CNNs regarding cross-channel correlation and classification complexity.
• To develop a computationally efficient and highly accurate model for automated glaucoma diagnosis.

Main Methods:

• Developed MCICNet, integrating convolution and involution operations in a cascade structure.
• Replaced traditional fully-connected layers with LightGBM for enhanced classification precision and generalization.
• Conducted extensive experiments on the LAG and ACRIMA datasets, comparing MCICNet against various CNN and transformer-based models.

Main Results:

• MCICNet achieved 95.6% accuracy on the LAG dataset and 96.2% on ACRIMA.
• Outperformed nine CNN architectures (AlexNet, MobileNetV2, SqueezeNet, ResNet18, GoogLeNet, DenseNet121, EfficientNetB0, ShuffleNet, VGG16) and three transformer models (ViT, MaxViT, SwinT).
• Demonstrated superior performance over its variant without involution (MCICNet-NoInvolution) with only 0.9 million parameters.

Conclusions:

• MCICNet offers significant improvements in glaucoma detection accuracy and efficiency.
• The combination of convolution, involution, and LightGBM effectively captures inter-channel dependencies and enhances classification.
• MCICNet presents an advanced, computationally efficient solution for automated glaucoma detection in medical imaging.