Skripsi
BILATERAL FILTERING, UNSHARP MASKING, DAN PIX2PIX GAN SEBAGAI UPAYA PENINGKATAN KUALITAS DAN AUGMENTASI CITRA RETINA PEMBULUH DARAH
Analysis of abnormalities in retinal blood vessels through deep learning methods is constrained by limited medical data, making augmentation techniques necessary to prevent overfitting. This study proposes the use of Pix2Pix Generative Adversarial Network (GAN) combined with transformation augmentation to increase the number of retinal images and generate synthetic images based on the distribution patterns of the original data. As an effort to optimize augmentation performance, this study implements an image enhancement stage using a combination of Bilateral Filtering (BF) and Unsharp Masking (UM) to reduce noise and emphasize contrast details in retinal blood vessels. This study uses the LES-AV dataset, consisting of 22 retinal images along with ground truth images of arterial and venous. The image enhancement performance using the BF and UM methods achieves the best results compared to other image enhancement method sequences. Evaluation of image enhancement performance resulted in a Root Mean Squared Error (RMSE) value of 2.31, Peak Signal-to-Noise Ratio (PSNR) of 40.89, and Structural Similarity Index Measure (SSIM) of 0.95. Implementing BF and UM enhanced images as input in the Pix2Pix GAN augmentation technique was able to improve augmentation performance compared to Pix2Pix GAN without image enhancement. Evaluation of Pix2Pix GAN performance yielded a PSNR of 43.19, SSIM of 0.94, and Frechet Inception Distance (FID) of 0.56. The augmentation results produced 506 images from transformation augmentation and then 1,012 images from Pix2Pix GAN augmentation.
No other version available