Quality Enhancement of Breast DCE-MRI Images Via Convolutional Autoencoders

Abstract

Magnetic Resonance Imaging (MRI) is a technique used in medicine to visualize interior body structures in order to diagnose a variety of disorders, including breast cancer. For its diagnosis, MRI is employed along with a chemical compound called contrast agent to enhance the visualization of tumoral structures in tissue. This examination is usually called dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). However, the resolution of the images or the MRI equipment measured in Teslas (T), is a vital parameter for correctly visualizing inside structures, but this feature increases their prices dramatically, making them inaccessible in some situations. As a result, scientific research to improve image quality without spending additional costs or equipment has lately emerged. This work proposes the use of a deep learning strategy to improve the quality of medical images obtained in DCE-MRI. Images from the Qin breast dataset are synthetically degraded by adding noise and applying a spatial transformation (size reduction). The original image is then reconstructed from the degraded ones by using a Convolutional Autoencoder (CAE) to minimize the difference between them. The results of computational similarity metrics such as mean squared error (MSE), peak signal to noise ratio (PSNR), and structural similarity index measure (SSIM) revealed an effective correspondence and visualization between the synthesized and original images at a broad level (full image) and in specific regions of interest (breast tissue).