Optimized Resolution-Oriented Many-to-One Intensity Standardization Method for Magnetic Resonance Images

With the development of big data, Radiomics and deep-learning methods based on magnetic resonance (MR) images, it is necessary to conduct large databases containing MR images from multiple centers. Having huge intensity distribution differences among images reduced or even eliminated, robust computer-aided diagnosis models could be established. Therefore, an optimized intensity standardization model is proposed. The network structure, loss function, and data input strategy were optimized to better avoid the image resolution loss during transformation. The experimental dataset was obtained from five MR scanners located in four hospitals and was divided into nine groups based on the imaging parameters, during which 9152 MR images from 499 participants were collected. Experiments show the superiority of the proposed method to the previously proposed unified model in resolution metrics including the peak signal-to-noise ratio, structural similarity, visual information fidelity, universal quality index, and image fidelity criterion. Another experiment further shows the advantage of the proposed method in increasing the effectiveness of following computer-aided diagnosis models by better preservation of MR image details. Moreover, the advantage over conventional standardization methods are also shown. Thus, MR images from different centers can be standardized using the proposed method, which will facilitate numerous data-driven medical imaging studies.