Advanced AI-Driven Clinical Decision Support for Retinal Disease Screening with Ultra-Wide-Field Fundus Imaging
NEGAR KHALAF1 , Hossein Parsaei 1 *, Tahereh Mahmoudi1 , Masoumeh Masoumpour 2 , Hajar Farvardin2
- Department of Medical Physics and Engineering, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran.
- Poostchi Ophthalmology Research Center, Department of Ophthalmology, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran
Abstract: Retinal diseases pose a significant threat to visual health, with early detection crucial for preventing irreversible vision loss. Despite advancements in imaging technology, many regions still lack access to specialized ophthalmic care, leading to delays in diagnosis. Ultra-wide-field (UWF) imaging, including color fundus and autofluorescence modalities, provides comprehensive retinal views, enhancing disease detection. This study aims to develop an AI-based system for screening retinal diseases using UWF imaging, leveraging deep learning (DL) techniques to classify retinal images as normal or abnormal, offering a scalable and efficient solution for clinical decision support.
Methods: The study utilized a dataset of 10,440 UWF color and autofluorescence images from 2,610 patients, collected over a 15-month period at Dr. Farvardin’s Eye Clinic. These images were captured using the OPTOS P200DTx/A10650 device and were carefully curated to remove artifacts such as operator interference or eyelid obstruction, ensuring high-quality data for training the AI models. Convolutional neural networks (CNNs), including DenseNet121, ResNet18, and InceptionResNetV2, were applied to classify images into normal or abnormal categories. Transfer learning was used to fine-tune pre-trained models for retinal disease detection. A combined decision-making approach was employed to optimize model performance by integrating predictions from multiple networks. Models were evaluated based on accuracy, sensitivity, and specificity.
Results: The CNN models achieved significant classification performance, with the 4-channel input model, combining UWF color and autofluorescence images(as 4-channel), yielding superior accuracy compared to 3-channel models. DenseNet121 demonstrated the highest performance in detecting retinal abnormalities, with a promising AUROC. The combined decision-making approach achieved the best performance overall, with a sensitivity of 90%, specificity of 90%, and an accuracy of 89%. The system successfully classified retinal images, facilitating early detection of pathological changes, including diabetic retinopathy, macular degeneration, and other retinal diseases.
Conclusion: The integration of deep learning with UWF imaging offers a promising tool for improving retinal disease screening, especially in underserved areas with limited access to ophthalmologists. This AI-driven approach can significantly enhance diagnostic accuracy and provide an efficient, scalable solution for the timely detection and management of retinal conditions.
