A robust intensity distribution alignment for harmonization of T1w intensity values

Variations in image intensities between magnetic resonance imaging (MRI) acquisitions affect the subsequent image processing and its derived outcomes. Therefore, it is necessary to normalize images of different scanners/acquisitions, especially for longitudinal studies where a change of scanner or pulse sequence often happens. Here, we propose a robust intensity distribution alignment (RIDA) method to remove between-scan effects. The method is based on MRI T1w images acquired in close succession and robustly aligns two cumulative distribution functions (CDF) of voxel intensities to improve image-derived outcomes of a range of subcortical brain structures with different acquisition parameters. We compare RIDA with the other image harmonization methods: mica and RAVEL. We study three intra-scanner and three inter-scanner protocol variations among the same 20 participants scanned with Siemens 1.5T Avanto, 3T Skyra, and 3T Prisma scanners on the same day and use image-derived volumetric outputs from the Sequence Adaptive Multimodal Segmentation (SAMSEG) method. We find that CDF-based intensity harmonization (mica and RIDA) significantly reduces intensity differences, improves consistency in volume quantification, and increases spatial overlap between two images acquired in close succession. The improvements are most considerable if the intensity normalization is based on subcortical structures only (RIDA), excluding cortical regions, instead of the whole brain. However, the effect of the corrections varies considerably as a function of the compared scanners and sequences. In conclusion, the RIDA scaneffect normalization improves the consistency of image-derived measures, but its performance depends on several factors.