Harmonizing data acquisition, reconstruction, and analysis for reproducible, cross-vendor, open source MRI

Abstract In this 5-year R01 project entitled “Harmonizing data acquisition, reconstruction, and analysis for reproducible, cross-vendor, open-source MRI,” we address the significant barriers to scientific progress due to the large inter- scanner variability (often more than 10-20%) present in multi-site MRI data which substantially diminishes the power of neuroimaging studies to detect subtle pathologies in neuropsychiatric disorders. Inter-scanner biases are a result of differences in implementation of closed-source product sequences (e.g., gradient and radiofrequency pulse shapes and timing), the choice of reconstruction algorithms, as well as variations inherent to the scanner hardware (e.g., gradient strength). Another major challenge is the significant barrier to develop new sequences for each vendor separately. This inhibits the translation of new MRI technologies to research laboratories, as vendor-specific sequence development environments are closed-source, proprietary, and suffer from a steep learning curve. In this project, we address these challenges by proposing an “end-to-end” harmonization framework. We propose to develop and disseminate a single open-source vendor-neutral MRI pulse sequence development environment containing both standard MRI protocols (e.g., T1-weighted, T2-weighted, and diffusion MRI) and cutting-edge quantitative acquisitions (T1, T2, T2*, and quantitative susceptibility maps (QSM)), a unified image reconstruction framework, and novel algorithms for post-acquisition data harmonization to enable multi-site reproducible research and mitigate inter-scanner variability and bias. Our quantitative MRI acquisitions will be efficient (5 min as opposed to more than 15 min) and also comprise of fast, distortion-free diffusion MRI sequences. The performance of standard contrast-weighted protocols and the accuracy of novel quantitative imaging sequences will be rigorously validated on phantoms and in-vivo data acquired from all major vendors (Siemens, Philips, GE). Further, we will develop and validate novel data harmonization algorithms that will remove any remaining scanner-induced discrepancies in the data due to hardware differences. One of the goals of this project is to reduce inter-scanner variability to the level of those observed within-scanner. The technical developments proposed in this grant will dramatically increase reproducibility across sites and allow for seamless execution of multi-site neuroimaging studies. Thus, the increased statistical power of multi-site studies will facilitate detection of subtle changes in neuropsychiatric disorders. Our open-source first-of-its-kind platform will also accelerate cross-vendor sequence development and enable immediate translation of new sequences into research studies (which currently takes several years).