Högskolan i Skövde

Glaucoma is an optic neuropathy characterised by irreparable damage in the form of gradual vision loss. As one of the leading causes of blindness worldwide, research into preventative healthcare such as accessible, early-stage screening mechanisms is rapidly growing in importance. Automated deep learning methods integrated with retinal fundus image datasets hold promise however balancing model capacity against calibration and resource constraints remains a major challenge within digital healthcare. 

This study systematically evaluated trade-offs between diagnostic performance, structural complexity and clinical calibration across convolutional neural networks (CNNs) to identify optimal frameworks for mobile health glaucoma screening. Five distinct architectures- EfficientNet B0, EfficientNet-B7, Efficient-V2-M, VGG16, and ResNet50- were optimised across a hyperparameter grid search using the REFUGE2 dataset and results evaluated were based on the optimal configuration for each architecture. 

Due to the dataset’s class imbalance (80% healthy), clinical viability was evaluated at the standard 0.5 decision boundary using F1 Score and classifier decision-boundary calibration was profiled using confidence strip plots. Continuous prediction probabilities from unseen test sets were extracted for each architecture to calculate Receiver Operator Characteristic (ROC) curves. Structural parameter weights were mapped against diagnostic performance to outline the experimental efficiency frontier. 

The study revealed that while high-capacity networks like EfficientNet-B7 (~64M parameters) can achieve isolated, peak theoretical metrics (highest AUC=0.8556), they can fall short at the clinical decision boundary (F1 Score = 0.5674). 

Conversely, the highly compact EfficientNet-B0 (~4M parameters), established a superior efficiency frontier and delivered a predictable, stable performance consisting of AUC=0.837, the highest overall F1 Score (0.643), and B0 yielded more polarised, confident predictions with a low False Positive Rate of ~5.3%. VGG16 and ResNet50 delivered poor performances despite increased parameter count compared to the lean, EfficientNet-B0 model with smeared predictions across the decision boundary, demonstrating higher diagnostic uncertainty. 

This study demonstrates that architectural scale does not guarantee clinical reliability on finite and specialised, biological datasets. Out of the evaluated models, EfficientNet-B0 is established as the definitive choice for clinical translation into low-resource settings due to its compact parameter profile and highly confident calibration.