The Performance of two Segmentation Convolutional Neural Networks, U-Net and SegNet, for the Segmentation of Cystoid Macular Edema
Abstract:
Background: Screening is the primary method used to detect cystoid macular edema in conditions that threaten vision, such as diabetic retinopathy and age-related macular degeneration. However, the interpretation of imaging data relies heavily on the expertise of ophthalmologists, leading to subjective factors. The integration of computer-aided diagnostics as a secondary tool has helped to decrease diagnostic uncertainties among doctors. Therefore, the refinement of these systems, particularly in terms of more precise segmentation and disease detection, is crucial. In this study, two well-known convolutional neural networks (CNNs), U-Net and SegNet, were utilized for segmenting intraretinal fluid in optical coherence tomography (OCT) images for diagnosing cystoid macular edema (CME). Determining the most suitable architecture for this task can alleviate the burden on ophthalmologists during CME screenings and triage.
Methods: CNNs are employed in segmentation to categorize individual pixels in an image based on self-learned weights. In this study, we utilized 80 OCT images of cystoid macular edema (CME) along with their respective masks obtained from the Liaquat University of Medical and Health Sciences dataset. We evaluated the efficacy of automatic segmentation conducted by U-Net and SegNet with various configurations in the task of intraretinal fluid segmentation relevant to CME diagnosis.
Results: Out of the two suggested architectures, U-Net demonstrated superior performance, achieving an accuracy of 0.9984, a dice coefficient of 0.8478, and requiring less training time compared to SegNet. In contrast, SegNet achieved an accuracy of 0.9961 and a dice score of 0.5502.
Conclusion: Depending on their setup, both networks can accurately segment OCT images of CME, with U-Net typically demonstrating quicker training times and higher accuracy levels.
