Harmonization And Optimization For Multicenter Diffusion MRI

Over the last decade,numerous multicenter studies of magnetic resonance imaging(MRI)have been performed for discovering the connection patterns of brain organizations and mechanisms of cognitive functions in normal,developing,and aging populations,as well as in various diseases.Diffusion MRI has become one of the main imaging modalities in these multicenter projects recently due to its capability for quantitative measurements of brain microstructure.In the acquisition of diffusion imaging,the multi-shell scheme is gradually popularized for being compatible with various diffusion models and post-processing algorithms.As a premise,the reproducibility and reliability of images are the key to evaluate data consistencies within one center and among centers.Currently,the diffusion tensor model with single-shell scheme is mostly used for evaluation of repeatability of multicenter diffusion data,however,the evaluation for other advanced diffusion models is less.Therefore,it is required to evaluate the reproducibility of multi-shell data on higher-order diffusion-derived measures.Moreover,the data consistency is reduced due to the varied acquisition parameters with different MRI scanners.Appropriate harmonization methods are demanded to improve data reliability and ensure that the differences among centers would not affect the quantitative analysis.This work was conducted to meet the above requirements.In the first part,the multicenter reproducibility of multi-shell data was evaluated.After a well-controlled experiment using the same type of MRI hardware and the same protocols on travelling subjects,the whole brain tractography was reconstructed using a high-order diffusion model.Later,from the local-scaled track density and the distal-scaled structural connectome matrix,it was found that the reproducibility of intra-center was still slightly higher than that of inter-center.And the results also showed that compared to single-shell,the multi-shell model could produce higher reproducibility and was more tolerance to image noise.In the second part of the work,a diffusion harmonization framework was designed based on deep learning network to harmonize the diffusion kurtosis imaging(DKI)metrics among MRI scanners to improve data consistency.By setting one scanner with higher imaging quality as reference,the neural network could be trained in other scanners,the results showed that the data reliability of the DKI metrics among all scanners were improved,and the data validity in each scanner was also increased.In addition,compared with other methods,harmonized data using high-efficient deep learning-based method was highest in data validity.The third part of the work was for the quality assessment of diffusion data collected in multiple centers.The quality of diffusion images collected from volunteers were first measured.An automated process for daily MRI quality control was then developed and deployed,which could facilitate the longitudinal monitoring of scanner status.Finally,for fixed postmortem brain tissue,the acquisition scheme was optimized on scanners with two different field strengths,resulting in improved quality of structural and diffusion images.