An evaluation of AI-based methods for papilledema detection in retinal fundus images

Abstract

The complexities inherent in diagnosing papilledema, particularly within the realm of neuro-ophthalmology, emphasize the pressing need for sophisticated diagnostic methodologies. This study highlights the application of novel models tailored explicitly for papilledema detection, distinguishing it from pseudo-papilledema and normal cases, through the strategic utilization of deep learning frameworks, specifically convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Leveraging hierarchical feature extraction from retinal images, the multi-paths CNN model successfully identifies crucial indicators of papilledema, while the cascaded model, integrating ResNet-50 and long short-term memory (LSTM), effectively captures sequential features. Leveraging a meticulously curated dataset comprising 18,258 fundus images, the presented models exhibit exceptional performance, with the multi-paths CNN achieving an accuracy of 99.97 %, and the LSTM model demonstrating an accuracy of 99.81 %. Comparative analysis showcases the unparalleled efficacy of the models, underscoring their potential in clinical diagnostics. Notably, they demonstrate robustness in occlusion sensitivity tests, highlighting their resilience in scenarios involving obscured image components. This pioneering study represents a significant milestone in papilledema detection, with the promise of advancing patient outcomes and streamlining healthcare practices. The proposed deep learning models not only offer precise diagnoses but also hold the potential to automate elements of the diagnostic workflow, alleviating the workload of healthcare professionals and enhancing overall patient care outcomes.