Comparison Stability And Repeatability Between ICA Algorithms Using ICASSO

1. BackgroundICA is a data-driven method that separates mixed observations to sources that are in highest statistically independency.In 1998, McKeown et al introduced spatial ICA for fMRI analysis. After getting more acceptances, ICA was developed in many aspects (e.g. subject/group level, back-reconstruction, algorithms, the optimum number of estimated components, etc.) and various scenarios (e.g. temporal concatenation or tensorial ICA). In early observations, Hui et al (2004) demonstrated that GICA has better performance due to the highestspatial detection power and relatively more accurate estimation of time-courses. ICA has many algorithms (INFOMAX, FAST, ERICA etc.) and some of these algorithms results depend on their parameter setting and initial starting points. So repeatability and stability of the optimization problem are unneglectable issues. Himberg et al (2004) performed ICASSO to validate for comparing ICA algorithms in terms of reliability andstability. Therefore, this study investigates the reliability of ICA estimates using ICASSO. 2. MethodICASSO iterates ICA many times and clusters all results. The basic idea is that a tight cluster of estimates is considered to be a candidate for including a good estimate. Our study is evaluated on ADHD 200 dataset (Oregon Health and Science University). To access the reliable results, raw data was first pre-processed. All the cleaned data were analyzed by GIFT toolbox and tested by 4 GICA algorithms(FAST, INFOMAX, JADE,COMBI). Each algorithm was iterated 30 times and estimated 30 ICs in each iteration. After that, All results were clustered in 30 clusters. ICASSO measured cluster quality index (Iq) for each IC that says how much that cluster is integrated. Finally, average IC maps in each cluster were correlated with resting state networks (RSNs) template for goal of additional validation.3. Results After clustering, each cluster was expected to have 30 members but it happened only for INFOMAX. Best Iq averages belong to INFOMAX (0.98) and FAST (0.91) respectively. Moreover, ICASSO projected ICs to the first two PCA directions in dimension reduction stage and showed Iq of each IC corresponds with the integrity of that IC; In fig.1 and fig.2, results of ICASSO for INFOMAX (a) and FAST (b) are shown. The results of spatialcorrelation with RSNs template were almost equal for all algorithms but for INFOMAX we observed higher correlations (e.g. 26th IC correlated 91.4. Conclusions Stability and repeatability of the ICs were investigated in 4 different algorithms; Our finding in the analyzed dataset verified several previous studies like Schpf et al (2010) and Soldati et al (2013) that shows INFOMAX and FAST have the best performance respectively. Based on our results, we suggest to utilize these two mentioned algorithms as recent studies took advantage of them e.g. Dash et al (2018)