Detection of keratoconus Diseases using deep Learning
Authors:
AKM Enzam-Ul Haque,
Golam Rabbany,
Md. Siam
Abstract:
One of the most serious corneal disorders, keratoconus is difficult to diagnose in its early stages and can result in blindness. This illness, which often appears in the second decade of life, affects people of all sexes and races. Convolutional neural networks (CNNs), one of the deep learning approaches, have recently come to light as particularly promising tools for the accurate and timely diagn…
▽ More
One of the most serious corneal disorders, keratoconus is difficult to diagnose in its early stages and can result in blindness. This illness, which often appears in the second decade of life, affects people of all sexes and races. Convolutional neural networks (CNNs), one of the deep learning approaches, have recently come to light as particularly promising tools for the accurate and timely diagnosis of keratoconus. The purpose of this study was to evaluate how well different D-CNN models identified keratoconus-related diseases. To be more precise, we compared five different CNN-based deep learning architectures (DenseNet201, InceptionV3, MobileNetV2, VGG19, Xception). In our comprehensive experimental analysis, the DenseNet201-based model performed very well in keratoconus disease identification in our extensive experimental research. This model outperformed its D-CNN equivalents, with an astounding accuracy rate of 89.14% in three crucial classes: Keratoconus, Normal, and Suspect. The results demonstrate not only the stability and robustness of the model but also its practical usefulness in real-world applications for accurate and dependable keratoconus identification. In addition, D-CNN DenseNet201 performs extraordinarily well in terms of precision, recall rates, and F1 scores in addition to accuracy. These measures validate the model's usefulness as an effective diagnostic tool by highlighting its capacity to reliably detect instances of keratoconus and to reduce false positives and negatives.
△ Less
Submitted 3 November, 2023;
originally announced November 2023.
