FS-Net: Medical image denoising via local receptive field smoothing network

Noise is one of the most challenging tasks in medical images which is happening during transmission and acquisition, and as a result, it complicates a subsequent medical image analysis task. This paper proposes a novel local receptive field smoothing network (FS-Net) for medical image denoising by introducing smoothing priors of receptive field of intermediate CNN layers. First, for a receptive field of a CNN, each of its values smoothing properties are retained by weighting their local neighborhoods (i.e., after smoothing, each new neurons are the weighted average of its neighborhoods). The weights are calculated using the similarity of their corresponding receptive field (i.e., the similarity between the receptive field used to calculate a receptive fields value and the receptive field used to calculate their corresponding neighborhoods) and this keeps the smoothing properties. Second, by applying this smoothing prior on receptive fields of intermediate CNN layers, an end to end FSNet which directly approximate the latent clean images from its noisy version is designed. The receptive fields smoothing helps to train deeper architecture and improves the reconstructed image quality. Experimental results on Breast Mammography images reveal that the proposed method outperforms other states of the art denoising methods.