From Noise Reduction to Disease Detection: An Evaluation of BM3D and Autoencoder on OCT Images with ROI Segmentation

Digital device usage and aging can lead to eye issues. Optical Coherence Tomography (OCT) scans diagnose eye conditions. However, noise can cause problems in OCT images for professionals and ordinary individuals to diagnose. In this context, we evaluated two noise-denoising techniques Autoencoder (AE) and BM3D (Block-matching and 3D filtering). Our proposed AE outperformed BM3D with a PSNR of 32.47 and a denoising time of 0.01 seconds. We also proposed a deep Convolutional Neural Network (CNN) method with transfer learning using a pre-trained architecture and a Fully Connected (FC) neural network as a classifier. Our proposed DenseNet121 model on the AE denoised dataset achieved 100% accuracy during the blind test with 0.0067 seconds prediction time for an image, showing its competitiveness against existing works.