Detection of optic nerve hypoplasia disease using deep learning

The accurate localization and evaluation of the optic disc position in human retinal images is critical for the early diagnosis of congenital anomalies such as optic nerve hypoplasia (ONH). This doctoral research presents a novel deep learning-based system utilizing convolutional neural networks for the early detection of ONH. The precise localization, measurement, and evaluation of the optic disc and fovea in retinal images are essential for ONH diagnosis. This automated diagnostic approach, proposed for the first time for ONH, leverages a deep learning model that combines a ResNet-18 pre-trained encoder with the U-Net architecture. The proposed system overcomes challenges associated with traditional methods, such as variations in illumination, blurred boundaries, and vascular occlusions, through gradient descent optimization and multi-scale feature extraction. The developed model has undergone extensive testing on international datasets, including Messidor, IDRID, DIARETDB1, HRF, DRIVE, and APTOS, achieving high performance in both optic disc and fovea segmentation. To evaluate its clinical applicability, an original dataset named ONH-NET was created using retinal images collected from the Department of Ophthalmology at Duzce University. The system's performance on real clinical cases was thoroughly validated. The most significant contribution of this study is achieving %99,58 accuracy in ONH diagnosis by automatically calculating the ratios of optic disc diameter to macular center distance. The system offers clinicians a fast, objective, and reliable decision support mechanism, eliminating the need for manual measurements and calculations. In addition to optic disc and fovea localization, the automatic determination of macular boundaries and critical morphometric measurements for ONH further distinguishes this study from similar works in the literature. By integrating innovative approaches in retinal imaging and deep learning, this doctoral thesis provides a practical solution for clinical applications and makes significant contributions to the field of ophthalmology.