ISSN: 2578-465X
Authors: Carolyn Yu Tung Wong
Background: Screening is the predominant strategy for early detection of ROP. However, image analysis is dependent on the experience of ophthalmologists, which introduces subjective variables into the interpretation of data. Computer-aided diagnostics has recently become more widely employed as a second-reader tool, reducing doctors' diagnostic uncertainties. Based on this, the development of such systems is critical, and it may be done by refining its processes, namely, more accurate segmentation and disease detection. Given this circumstance, the current study presents the use of two well-known convolutional neural networks (CNNs) developed for segmentation, U-Net and SegNet, to segment blood vessels (BV), optic discs (OD), and demarcation lines/ridges in ROP fundus images. Defining which architecture is most suited for such a task can assist in reducing the stress on ophthalmologists for ROP screenings and triage. Methods: CNNs are used in segmentation to classify each pixel in an image using self-trained weights. Using three data subsets, each containing 50 RetCam images of ROP and their corresponding masks regarding the BV, OD, or ridge of interest, retrieved from the only publicly available ROP segmentation dataset, i.e. HVDROPDB, we compared the automatic segmentation performed by the U-Net and SegNet using different configurations and in different segmentation tasks related to ROP diagnosis. Results: Among the two proposed architectures, U-Net obtained better overall results in all segmentation tasks, obtaining accuracies of 0.933-0.8751, dice coefficients of 0.49-0.648, and took less training time than the SegNet, which achieved accuracies of 0.933-0.80 and dice scores of 0.40-0.648. Conclusion: The two networks can segment RetCam images of ROP with useful accuracy depending on their configuration, with U-Net being generally faster to train and more accurate
Keywords: Artificial Intelligence; Image Segmentation; Retinopathy of Prematurity; Vasoproliferative
