Deep Learning Networks with p-norm Loss Layers for Spatial Resolution Enhancement of 3D Medical Images

Abstract

0000−0001−6519−1213] , Ezequiel López-Rubio 1[0000−0001−8231−5687] , Núria Roé-Vellvé 2 , and Miguel A. Molina-Cabello 1[0000−0002−8929−6017]Abstract. Nowadays, obtaining high-quality magnetic resonance (MR) images is a complex problem due to several acquisition factors, but is crucial in order to perform good diagnostics. The enhancement of the resolution is a typical procedure applied after the image generation. Stateof-the-art works gather a large variety of methods for super-resolution (SR), among which deep learning has become very popular during the last years. Most of the SR deep-learning methods are based on the minimization of the residuals by the use of Euclidean loss layers. In this paper, we propose an SR model based on the use of a p-norm loss layer to improve the learning process and obtain a better high-resolution (HR) image. This method was implemented using a three-dimensional convolutional neural network (CNN), and tested for several norms in order to determine the most robust fit. The proposed methodology was trained and tested with sets of MR structural T1-weighted images and showed better outcomes quantitatively, in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), and the restored and the calculated residual images showed better CNN outputs.

Artificial intelligence within the interplay between natural and artificial computation: Advances in data science, trends and applications

“…Most of these CNNs are based on the minimization of the residuals using the squared Euclidean cost function. [120] proposed a novel optimization algorithm for CNNs based on the p-norm, where p is the exponent of the norm, which can reduce the effect of outliers and improve the convergence of the network. The use of values p < 2 reduces the influence of extreme values of the residual error, i.e.…”