Robust optic disc localization using fast RCNN and aggregate channel features (ACF) with generative adversarial network (GAN) synthesized fundus image / Nurhakimah Abd Aziz

Recent advancements in artificial intelligence (AI) hold promise in addressing the challenge of blindness, particularly through accurate and non-invasive early detection of Diabetic Retinopathy (DR). Among the prominent symptoms of DR, exudates play a crucial role. Detecting these symptoms in fundus images is intricate due to the visual resemblance between the optic disc (OD) and exudates. Hence, discerning the OD before detecting exudates bears significance. This study aims to develop OD detection systems using both Deep Learning (DL) and non-deep learning (non-DL) methods. The proposed model was trained and evaluated using various publicly available fundus image datasets, including Kaggle, DIARETDB, DRIMDB, and Messidor. Preprocessing techniques were applied to enhance the image quality, involving color normalization through the Reinhard method, OD labeling, image resizing, and dataset expansion. To ensure the robustness of DL-based techniques, which rely on extensive examples, a Generative Adversarial Network (GAN) based on the Lightweight GAN architecture was introduced to synthesize fundus samples for AI training. Subsequently, the DL model Fast Region-based CNN (Fast RCNN) with transfer learning and the nonDL model, Aggregate Channel Features (ACF), were trained and assessed using the processed and GAN-generated datasets. The trained GAN network proficiently generated high-resolution samples of normal and diseased fundus images. The OD detection achieved an average confidence score above 90% for training, testing, and validation datasets for both methods. Performance evaluation indicated that Fast RCNN exhibited the highest average precision (AP), specifically 80.79%, 84.67%, and 90.22% for the test, validation, and GAN datasets, respectively. Meanwhile, the ACF method achieved AP of 75.3%, 79.1 and 80.12% for the respective datasets. Although Fast RCNN outperformed ACF across all dataset categories, it's worth highlighting that ACF's detector still delivered a commendable performance, with minimum rate of correct detection achieved 75% and maximum of correct detection approximately 90%. These outcomes highlight the effectiveness of the proposed approaches in handling multi-sourced datasets characterized by non-standard colour, illumination, quality, location, and acquisition devices.