Deep Learning Autoencoder for denoising Medical Images

Abstract

Medical image examination is crucial for accurate diagnosis and treatment plan of medical disorders in the human body. Medical images are corrupted by noise, which obscures important details, leading to misinterpretation and wrong diagnosis of the images. Image denoising methods are employed to eliminate noise with a view to improving image quality and preserving its essential features. This work utilizes Deep Learning Autoencoder (DLA) for removing noise in medical images, and its performance is compared with U-Net (segmentation model), which, to the best of our knowledge, has not been presented in this manner in the literature. The metrics utilized for comparison are image quality, Peak Signal to Noise Ratio (PSNR), Mean Squared Error (MSE), and Structural Similarity Index Measure (SSIM). It is found that DLA outperforms U-Net, producing better image quality and, in addition, has lower MSE, higher PSNR and SSIM. For instance, the MSE, PSNR, and SSIM of DLA when used to denoise noisy lung cancer images are 0.0226, 16.584dB, and 0.3845, respectively, while those of U-Net are 0.2487, 6.118dB, and 0.04760. In addition, it is found that the performance of DLA surpasses that reported for the state-of-the-art models in the literature.