Bramsh Qamar Chandio^{1}, Jaroslaw Harezlak^{2}, Serge Koudoro ^{1}, David Reagan ^{3}, and Eleftherios Garyfallidis^{1}

Bundle Analytics promises fast, robust, and flexible computational and statistical analyses for tractometric studies on clinical data. It uses information from both tractometry, and anatomy to analyze the extracted fiber bundles from challenging clinical datasets. It uses streamline-based efficient algorithms to register and extract fiber bundles from a tractogram, and applies linear mixed models in the extracted bundles to find significant differences at specific locations of the bundles across groups. Finally, the method does not require training, an important advantage over deep learning methods.

There is an exciting explosion of methods that analyze tractograms and combine them with anatomical information. However, building such pipelines needs careful consideration especially for clinical applications. In this paper we present Bundle Analytics, a framework that can precisely extract bundles and perform statistical analysis across populations. The input of the framework are tractograms and anatomy files. This framework is robust, fast, flexible, and applicable on every kind of data including hard to process clinical data.

Bundles analytics uses the advantages of streamline-based analysis to solve hard clinical analysis problems. It takes a target tractogram and registers it to a model tractogram using Streamline-based Linear Registration algorithm^{1 }. Model bundle is used to extract bundle from registered tractogram using a new version of RecoBundles^{ 2} . This refined version of RecoBundles uses output from standard RecoBundles as new model bundle and extracts more streamlines from tractogram. It allows us to reach accurate streamlines as new model bundle is already part of tractogram. Finally, we register and autoprune this new extracted bundle with original model bundle. This method is useful when we are dealing with noisy data or when we want to extract small bundles from tractograms. See Fig.1 for a visual description of the workflow.

For the statistical part of our framework, we segment the bundles into segments using the model bundle centroids along their lengths in common space. We call this the assignment map. For creating this assignment map, we calculate euclidean distances between every point on streamline to all the model bundle centroids and assign it to the nearest segment centroid similar to Cousineau et al. ^{7}. Assignments are created in common space, which ensures that segment index corresponds to the same centroid across all populations. Fig.2 Shows visual process of creating assignment maps on bundle in common space.

In the final step, we take bundles back into native space and use information from anatomy to perform statistical analysis on bundles in native space. In this abstract, we are using DTI measures, FA, AD, MD, and RD along with CSA and CSD peak values (generated from DIPY^{6}). we apply linear mixed models on these metrics for statistical analysis. Fig.3 shows the different metrics visualized on fiber bundle in native space.

[1] Garyfallidis et al. Robust and efficient linear registration of white-matter fascicles in the space of streamlines, NeuroImage, 117, 124--140, 2015

[2] Garyfallidis et al. Recognition of white matter bundles using local and global streamline-based registration and clustering, Neuroimage, 2017.

[3] The Michael J. Fox Foundation for Parkinson's Research, New York, NY

[4] Yeh et al. NTU-90: a high angular resolution brain atlas constructed by q-space diffeomorphic reconstruction. Neuroimage. 2011 Sep 1;58(1):91-9

[5] Ngattai Lam et al. TRAFIC: Fiber Tract Classification Using Deep Learning, NCBI, 2017

[6] Garyfallidis et al. Dipy, a library for the analysis of diffusion MRI data Front. Neuroinform., 8 (2014)

[7] Cousineau et al., A test-retest study on Parkinson's PPMI dataset yields statistically significant white matter fascicles, NeuroImage: Clinical, 2017

[8] Dayan et al., Profilometry: A new statistical framework for the characterization of white matter pathways, with application to multiple sclerosis, NCBI, 2016

[9] Yeatman et al., Tract Profiles of White Matter Properties: Automating Fiber-Tract Quantification, NCBI, 2012

[10] Wasserthal et al., TractSeg - Fast and accurate white matter tract segmentation, NeuroImage, 2018

[11] https://brainlife.io

Workflow used to register the target tractogram to the model tractogram using streamline-based linear registration, and extract a bundle from registered tractogram using refined version of Recobundles.

Assignment maps created in common space using model bundle centroids along the length in common space. Assignment map of a) extracted bundle is created using b) model bundle centroids (100 segments). c) We calculate euclidean distances from every point on streamline to all the centroids of model bundle and assigning points to nearest segments as shown in (d).

FA, AD, MD, RD, CSA peak values, CSD peak values metrics visualized on arcuate fasciculus in native space. All metrics normalized from 0-1 for

Summary of the group comparisons on the FA measure. Each solid line represents the smoothed “-log10(p-value)” from the group comparison for four considered tracks. Dotted line is plotted at the significance level equal to 0.01 and broken line at 0.05. Notice that there is a significant difference in FA and CSA (peak_values) are more significant than CSD (peak_values).