Abstract:
Keratoconus (KCN) is a progressive, non-inflammatory corneal disease that typically begins in early adulthood and leads to vision impairment due to symptoms like corneal thinning, anterior protrusion, and irregular astigmatism. Early detection is critical for managing and preventing the worsening of these symptoms. While various methods exist for diagnosing keratoconus, there remains a need for more accurate and efficient early detection techniques. The primary goal of this research is to improve the early detection of keratoconus from corneal morphogeometric images using advanced fine-tuned deep convolutional neural network (D-CNN) techniques. Previous studies have utilized CNN models for medical image analysis, but there is a gap in detection of keratoconus with high accuracy and efficiency. To address this gap, our research aims to provide a fine-tuned pre-trained D-CNN model for its early diagnosis which is crucial for preventing disease progression. This research has been used a merged dataset of seven mapping morphogeometric images, categorized into normal eyes, keratoconus, and suspected keratoconus. To prepare the images for analysis, we applied augmentation techniques such as GaussianBlur, LinearContrast, and AdditiveGaussianNoise to enhance image quality. We then trained the five fine-tuned deep learning based Convolutional Neural Network (CNN) models like MobileNetV2, InceptionV3, Xception, VGG19, and DenseNet201 techniques and then conducted experiments using the dataset. For KCN detection, the models achieved an AUC of 98% for the three-class problem for the augmented dataset. InceptionV3 provides better overall accuracy 89% and outperforms other CNN architectures. This model shows 4.6% improvement in testing accuracy over previous method on 4,011 (seven mapping morphogeometric) images dataset, also showing recall (0.89) and F1-scores (0.89) with minimum Type II error than other models. In future, we plan to build on the primary dataset and develop a reliable application using the Python FLASK web framework.