iPyrDAE: Image Pyramid-Based Denoising Autoencoder for Infrared Breast Images

Abstract

An early and accurate breast cancer diagnosis will reduce the death rate and improve survival chances. Thermography is a promising non-invasive early detection modality for breast cancer. Artificial intelligence-based classification systems are being used to classify thermographic infrared images. The success of these classification systems also depends on the quality of infrared images. However, the thermographic images acquired using digital infrared cameras are inevitably degraded by noise. In this paper, we propose novel denoising auto-encoders using image pyramids. The denoising auto-encoders are forced to learn the non-local noises by corrupting the input images in the image pyramid domain. We also propose a method to estimate the noise probability distribution in infrared images using statistical methods. The proposed denoising auto-encoder framework will learn better data representations from the corrupted images generated using image pyramids and estimated noise probability distribution. The Experimentation with four different infrared breast image data sets demonstrates robust representation learning by the agent showing promising improvements in the peak signal-to-noise ratio.