A Comparative Study on Deep Convolutional Neural Networks and Histogram Equalization Techniques for Glaucoma Detection From Fundus Images

AimsTo evaluate the performance of eight different convolutional neural network (CNN) models and histogram equalization techniques for glaucoma detection from fundus images. Material and MethodsThe study utilized the ACRIMA database, comprising 705 fundus images (396 glaucomatous and 309 normal). The CNN architectures evaluated include VGG-16, VGG-19, ResNet-50, ResNet-152, Inception v3, Xception, DenseNet-121, and EfficientNetB7. Two histogram-based preprocessing methods were applied: histogram equalization (HE) and contrast limited adaptive histogram equalization (CLAHE). The models were trained using supervised learning, with image augmentation techniques applied. Performance metrics such as accuracy, sensitivity, specificity, precision, negative predictive value, Dice Similarity Coefficient, and Area Under the Receiver Operating Characteristics Curve (AUC ROC) were used for evaluation. ResultsVGG-19 achieved the highest accuracy (97.9%) in the raw data setting, followed closely by VGG-16 (97.2%). ResNet-50 and ResNet-152 showed the highest specificity scores (98.4%). The sensitivity of the models ranged from 91.3% to 98.8%, with VGG-16 and VGG-19 demonstrating the highest values. CLAHE preprocessing resulted in improved performance, particularly for ResNet-152, which achieved an accuracy of 97.5% and sensitivity of 97.9%. The AUC ROC values varied from 0.937 to 0.996, with VGG-16 having the highest value (0.998). ConclusionThe study highlights the importance of selecting suitable CNN architectures and preprocessing techniques in developing effective glaucoma detection systems. While VGG-19 exhibited the highest accuracy with raw data, VGG-16 and ResNet-50 offered consistent performance across different preprocessing techniques, making them reliable options for clinical applications.