ARTIFICIAL SPOTS DETECTION IN FUNDUS IMAGES USING WAVELET-BASED U-NET

Deep learning algorithms help in detecting fundus disorders, including glaucoma, hypertension, and diabetic retinopathy at an early stage. Dust particles (artificial spots) may be added during the capturing of the fundus image, which can cause misdiagnosis as a disease. U-Net-based models have been used nowadays for semantic segmentation purposes. A wavelet features-based U-Net (WFU-Net) has been proposed to identify the non-standardized scanning-induced artificial spots. It creates a highly flexible feature fusion strategy by rethinking skip connections to aggregate the wavelet feature module (WFM). WFM is a feature fusion strategy, as skip connections are used in this paper, which helped WFU-Net to extract wavelet Haar transform features and fuse them to the decoder’s end. WFU-Net is a customized version of the Wide U-Net. This study evaluates prominent semantic segmentation techniques for artificial spot detection and compares those semantic segmentation techniques with WFU-Net. In terms of intersection over union (IOU), dice similarity coefficient (DSC), mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index matrix (SSIM) scores, the WFU-Net performs better. The IOU, DSC, PSNR, MAE, and SSIM of the WFU-Net are 0.8450, 0.9111, 0.0014, 31.39, and 0.9928, respectively.