MICCAI-CDMRI 2023 QuantConn Challenge Findings on Achieving Robust Quantitative Connectivity through Harmonized Preprocessing of Diffusion MRI

TL;DR

It is found that bundle surface area, fractional anisotropy, connectome assortativity, betweenness centrality, edge count, modularity, nodal strength, and participation coefficient measures are most biased by acquisition and that machine learning voxel-wise correction, RISH mapping, and NeSH methods effectively reduce these biases.

Abstract

White matter alterations are increasingly implicated in neurological diseases and their progression. International-scale studies use diffusion-weighted magnetic resonance imaging (DW-MRI) to qualitatively identify changes in white matter microstructure and connectivity. Yet, quantitative analysis of DW-MRI data is hindered by inconsistencies stemming from varying acquisition protocols. There is a pressing need to harmonize the preprocessing of DW-MRI datasets to ensure the derivation of robust quantitative diffusion metrics across acquisitions. In the MICCAI-CDMRI 2023 QuantConn challenge, participants were provided raw data from the same individuals collected on the same scanner but with two different acquisitions and tasked with preprocessing the DW-MRI to minimize acquisition differences while retaining biological variation. Submissions are evaluated on the reproducibility and comparability of cross-acquisition bundle-wise microstructure measures, bundle shape features, and connectomics. The key innovations of the QuantConn challenge are that (1) we assess bundles and tractography in the context of harmonization for the first time, (2) we assess connectomics in the context of harmonization for the first time, and (3) we have 10x additional subjects over prior harmonization challenge, MUSHAC and 100x over SuperMUDI. We find that bundle surface area, fractional anisotropy, connectome assortativity, betweenness centrality, edge count, modularity, nodal strength, and participation coefficient measures are most biased by acquisition and that machine learning voxel-wise correction, RISH mapping, and NeSH methods effectively reduce these biases. In addition, microstructure measures AD, MD, RD, bundle length, connectome density, efficiency, and path length are least biased by these acquisition differences.

Figures (7)

• Figure 1: We released 206 scans across two acquisitions, “A” (blue) and “B” (orange). Acquisition A was acquired with anisotropic resolution and 27 gradient directions. Acquisition B was acquired with isotropic resolution and 94 gradient directions. Participants altered the DW-MRI with the harmonization of their choosing. We then feed this harmonized data through a standard processing pipeline of tensor fitting, orientation distribution function (ODF) estimation, tractography, and tractometry to arrive at diffusion metrics of bundles microstructure and macrostructure and connectomics. Harmonization efficacy is determined by its capacity to minimize differences in these downstream diffusion measures.
• Figure 2: Using the harmonized data provided by participants, the full testing pipeline is as follows: tensor fitting, fODF estimation, whole brain tractography, bundle segmentation and tractometry, then connectomics, and finally complex network analysis. These processes result in three groups of analysis: complex network measures, bundle microstructure, and bundle macrostructure, which we evaluate the submissions on.
• Figure 3: Successful harmonization methods will reduce significant acquisition effects in these measures from the un-harmonized reference (“Ref”). Slashes indicate significant difference (p $\textless$ 0.05) in median between measures derived from acquisitions A and B. We compute Cohen’s D effect-size differences between connectomics measures from acquisitions A and B.
• Figure 4: We evaluate each submission on their ability to harmonize macrostructural features of 6 bundles. Successful harmonization will reduce significant acquisition effects in these features from the reference (“Ref”). We report normalized effect-size with Cohen’s D. Slashes indicate a significant difference (p $<$ 0.05) in median between features derived from acquisitions A and B.
• Figure 5: We compare the cross-acquisition (Acquisition A is blue, Acquisition B is orange) shape agreement of reconstructed bundles for one subject in the un-harmonized reference dataset and top performing harmonization technique (The Harmonizers 1) with the BUAN shape similarity metric. Shape similarity scores are reported in the upper right corner of each visualization, with $S$ being the BUAN shape similarity for reference data and $S$* the BUAN shape similarity for harmonized data. Bundle shape similarity score ranges from 0 to 1. The higher the value, the more similar the two bundles are in shape, and the lower value suggests low shape similarity.