Intensity standardization of multi-vendor MRI-a preliminary study

Abstract

Magnetic Resonance Imaging (MRI) has been used ubiquitously in medical fields for diagnosis, radiotherapy, or planning for surgery. MRI provides enhanced soft tissue contrast and high spatial resolution, resulting in better discriminating fat, water, muscle, and other soft tissues. However, the relationship between tissue intensity values in MR images is inconsistent, even though they are obtained with the same scanner and acquisition protocols and even within the same patients. Thus, this no tissue-specific intensity hinders accurate image processing (e.g., medical image segmentation and registration) and detection or classification of disease. Therefore, we designed the network to standardize the MR images acquired with multiple vendors and scan protocols. Our proposed model is a Generative Adversarial Network (GAN) in which the generator consists of two networks (Superresolution and style transfer), and the discriminator is a PatchGAN discriminator. We also designed the customized loss function using Radiomics, namely Radiomics Feature (RF) loss function. The histogram distance, image similarity metrics, and radiomics features are used for the evaluation metrics. We first compare the metric values between standard images and two synthesized images (Ny？l normalized (Histogram matching normalization) images and synthesized images via our proposed model). Additionally, the performances of other deep learning models are compared with our proposed model.