Multiscale Feature Attention Module Based Pyramid Network for Medical Digital Radiography Image Enhancement

Medical digital radiography (DR) is widely used in the clinical application. To deal with the problems of noise, edge blur, low contrast in DR images, we propose a multiscale feature attention module based pyramid enhancement network by training image blocks. The network is in the framework of a simplified U-Net, which reduces the computational load by reducing the convolution layer, and adopts Laplacian pyramid connection instead of concatenation operation to preserve the image boundary information. In addition, we embed a simple multiscale feature attention (SMFA) module between the encoder and decoder, which integrates the feature information of different scales precisely and makes the network have a stronger ability to perceive the local feature information. Our proposed algorithm is a network realization of Gauss-Laplacian pyramid decomposition with an attention module. Furthermore, we design a side feature loss function combined with mean square loss and absolute loss. We adopt batch normalization between convolution and activation operations to ensure information of all gray scale regions to be considered, which enhances the robustness of the network. We use LeakyReLu activation function and Sigmoid function in the previous layers and in the output layers respectively to preserve the negative information of multiscale details and to keep the gray scale region of the output images. Experiments with real data of different parts of human body validate the effectiveness of our algorithm, which shows that our proposed algorithm performs well on contrast enhancement, structure details preservation, and noise suppression. It has certain value of clinical application.