Deep Learning-based Glaucoma Diagnosis Techniques using Fundus Images

Abstract

In this thesis, we conducted research on glaucoma diagnosis in the field of ophthalmology and medicine using deep learning-based object detection and classification techniques with various retinal images. In the medical field, artificial intelligence diagnostic technologies based on deep learning are rapidly advancing. Most widely adopted deep learning-based diagnostic models to date are based on the common supervised learning (SL) paradigm. Supervised learning models require an adequate amount of training data for each class, and underfitting may occur in cases of insufficient data, making it challenging to ensure accurate diagnostic performance. The construction of deep learning training datasets in the medical field is challenging due to the obligation to protect personal medical information, making it difficult to acquire a sufficient amount of training data. Additionally, the majority of acquired data is from patients, leading to a relative scarcity of normal data. Therefore, not only conventional supervised learning techniques but also the utilization of various data fusion techniques are necessary for improving diagnostic performance. Furthermore, there is a demand for the development of algorithm models that can achieve sufficient learning and accurate diagnosis with limited amounts of training data. First, we proposed a deep learning algorithm model for glaucoma diagnosis. The proposed algorithm diagnoses the presence of glaucoma using Convolutional Neural Network (CNN) architecture. The data utilized included Ultra-Wide-Field (UWF) fundus imaging and True-color confocal scanner images, and their diagnostic capabilities were compared with the traditional Optical Coherence Tomography (OCT) parameter-based method. The experimental results confirmed the similarity between the diagnoses based on both types of retinal images obtained through deep learning and the conventional OCT parameter-based diagnoses. In conclusion, both deep learning-based UWF fundus imaging and True-color confocal scanner images can be effectively used for accurate glaucoma diagnosis. Second, we introduced Fusion by Convolution Network (FCN) and Fusion by Fully-connected Network (FFC), which extract and fuse features from various medical images. Among the nine types of data outputted from Swept-Source Optical Coherence Tomography (SS-OCT), the four most commonly used types were selected. After passing through CNN, the extracted features were fused to make a diagnosis of the presence of glaucoma and its progression. Experimental results showed higher diagnostic performance when utilizing a variety of image data compared to using only one type of image data. The diagnostic ability was superior to conventional OCT parameters. Lastly, we proposed Few-shot learning (FSL) as a technique for achieving sufficient learning and high accuracy in glaucoma diagnosis with a small amount of training data. The proposed FSL technique uses the ProtoNet algorithm model, which requires two learning processes: pre-training and training. In the pre-training stage, the model learns the characteristics of glaucomatous diseases using SS-OCT retinal images, while the training stage involves learning to distinguish between data features using Mini-Imagenet data. Diagnostic performance validation experiments were conducted using Wide-Field OCT Angiography (WF-OCTA) retinal images. The results showed that the conventional SL technique failed to learn with a small amount of WF-OCTA retinal image data, whereas the FSL technique demonstrated accurate diagnosis. Additionally, the diagnostic performance of the FSL technique was higher than that of the conventional OCT parameter-based approach.