ReAl-LiFE: Accelerating the Discovery of Individualized Brain Connectomes on GPUs
Authors: Sawan Kumar, Varsha Sreenivasan, Partha Talukdar, Franco Pestilli, Devarajan Sridharan630-638
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here, we introduce a GPU-based implementation of Li FE that achieves 50-100x speedups over conventional CPU-based implementations for connectome sizes of up to several million fibers. Linear classifiers, based on support vector machines, achieved over 80% accuracy in classifying AD patients from healthy controls based on their Re Al-Li FE pruned structural connectomes alone. |
| Researcher Affiliation | Academia | 1Computational and Data Sciences, Indian Institute of Science, Bangalore, India, 2Centre for Neuroscience, Indian Institute of Science, Bangalore, India, 3Computer Science and Automation, Indian Institute of Science, Bangalore, India, 4Psychological and Brain Sciences, Indiana University, Bloomington, USA |
| Pseudocode | Yes | Table 1: GPU Implementation of y = M times x(Φ, D, x) and Table 2: GPU Implementation of x = M transp y(Φ, D, y) |
| Open Source Code | Yes | We also provide an open source implementation of Re Al-Li FE1, along with software and hardware details needed for reproducing the results in this paper, to readily enable its application to other datasets. 1https://github.com/Sawan Kumar28/real-life |
| Open Datasets | Yes | Data from 90 individuals were drawn from the ADNI database (Mueller et al. 2005) comprising d MRI scans of patients with Alzheimer s Disease (AD; n=45, age: mean = 74.9; std = 8.5) and healthy, age-matched controls (NC; n=45, age: mean = 72.6; std = 5.6). We also confirmed these speedups on two other independent datasets acquired at other scanners. First, we optimized a d MRI dataset, H2. Dataset H data was already minimally preprocessed, and no additional proccessing was done. 2https://www.humanconnectome.org/study/hcp-youngadult/document/1200-subjects-data-release. Finally, we optimized a second d MRI dataset, S3 with d MRI scans already preprocessed based on a standard pipeline (Pestilli et al. 2014; Caiafa and Pestilli 2017). 3https://purl.stanford.edu/cs392kv3054 |
| Dataset Splits | Yes | We adopted the following approach: we performed tractography with data from one dataset and pruned the connectome by optimizing it, either with Li FE or Re Al-Li FE, on the same dataset. In each case, we generated a predicted diffusion signal from the fitted diffusion model, as described above (equation (2)). Next, we compared the root-mean-squared error (RMSE) between the predicted diffusion signal, in each case, and a second, independent, diffusion imaging dataset acquired from the same subject(s). We performed this cross-validation with data from two datasets: I and S (see Section 4.1). This procedure was repeated 150 times for random classifications of training and testing data, and classification accuracy was averaged across these runs. |
| Hardware Specification | Yes | Our baseline numbers are derived from the same configuration as Gugnani et al s Cluster A (single core Intel Xeon E5), enabling us to directly compare our speedup factors with theirs. |
| Software Dependencies | Yes | We developed a GPU-accelerated version of Li FE s SBB-NNLS optimization algorithm, with a CUDA implementation (Nvidia 2018). Whole brain probabilistic tractography was performed using the grey-matter-white-matter interface as a seed region. For each subject, anatomical (grey and white-matter) segmentations were obtained using Freesurfer v6 (Fischl et al. 2004). We then employed these connectivity measures as features in a support vector machine (SVM) classifier to test whether AD patients could be accurately distinguished from normal controls. For SVM classification, we used Matlab s fitclinear function, with soft margin, employing a linear kernel with C=1 (default values in Matlab). generated connectomes of various sizes ranging from 0.5 million to 2 million fibers with MRtrix v3.0 (Tournier et al. 2012). |
| Experiment Setup | Yes | We tested five different connectome sizes for 500 iterations of the optimization algorithm (Fig. 2A). For Re Al-Li FE, we tested several values of the regularization penalty parameter λ, on the fiber weights. For SVM classification, we used Matlab s fitclinear function, with soft margin, employing a linear kernel with C=1 (default values in Matlab). |