Compression and Denoising of Medical Images Using Autoencoders

In medical image analysis Image compression and denoising is an important processing steps for remote analytics. A number of algorithms are proposed in the literature with varying degrees of denoising performances. In this paper, a 3-layer autoencoder model to compress and denoise grayscale medical images is proposed. Gaussian noise with a noise factor of 0.5 is added and passes through 3 convolution and 3 max pool layers. The reverse process is made for image denoising. The proposed model learns to denoise data with a high PSNR and a high compression ratio after training. Most of the denoised images show over 30 dB PSNR with a 4:1 compression ratio. Images can be combined to increase sample size and overall denoising performance.