Brain extraction (a.k.a. skull stripping) is a fundamental step in the neuroimaging pipeline as it can affect the accuracy of downstream preprocess such as image registration, tissue classification, etc. Most brain extraction tools have been designed for and applied to human data and are often challenged by non-human primates (NHP) data. Amongst recent attempts to improve performance on NHP data, deep learning models appear to outperform the traditional tools. However, given the minimal sample size of most NHP studies and notable variations in data quality, the deep learning models are very rarely applied to multi-site samples in NHP imaging. To overcome this challenge, we used a transfer-learning framework that leverages a large human imaging dataset to pretrain a convolutional neural network (i.e. U-Net Model), and then transferred this to NHP data using a small NHP training sample. The resulting transfer-learning model converged faster and achieved more accurate performance than a similar U-Net Model trained exclusively on NHP samples. We improved the generalizability of the model by upgrading the transfer-learned model using additional training datasets from multiple research sites in the Primate Data-Exchange (PRIME-DE) consortium. Our final model outperformed brain extraction routines from popular MRI packages (AFNI, FSL, and FreeSurfer) across a heterogeneous sample from multiple sites in the PRIME-DE with less computational cost (20 s~10 min). We also demonstrated the transfer-learning process enables the macaque model to be updated for use with scans from chimpanzees, marmosets, and other mammals (e.g. pig). Our model, code, and the skull-stripped mask repository of 136 macaque monkeys are publicly available for unrestricted use by the neuroimaging community at https://github.com/HumanBrainED/NHP-BrainExtraction .

Compelling evidence suggests the need for more data per individual to reliably map the functional organization of the human connectome. As the notion that ‘more data is better’ emerges as a golden rule for functional connectomics, researchers find themselves grappling with the challenges of how to obtain the desired amounts of data per participant in a practical manner, particularly for retrospective data aggregation. Increasingly, the aggregation of data across all fMRI scans available for an individual is being viewed as a solution, regardless of scan condition (e.g., rest, task, movie). A number of open questions exist regarding the aggregation process and the impact of different decisions on the reliability of resultant aggregate data. We leveraged the availability of highly sampled test-retest datasets to systematically examine the impact of data aggregation strategies on the reliability of cortical functional connectomics. Specifically, we compared functional connectivity estimates derived after concatenating from: 1) multiple scans under the same state, 2) multiple scans under different states (i.e. hybrid or general functional connectivity), and 3) subsets of one long scan. We also varied connectivity processing (i.e. global signal regression, ICA-FIX, and task regression) and estimation procedures. When the total number of time points is equal, and the scan state held constant, concatenating multiple shorter scans had a clear advantage over a single long scan. However, this was not necessarily true when concatenating across different fMRI states (i.e. task conditions), where the reliability from the aggregate data varied across states. Concatenating fewer numbers of states that are more reliable tends to yield higher reliability. Our findings provide an overview of multiple dependencies of data concatenation that should be considered to optimize reliability in analysis of functional connectivity data.

Brain extraction (a.k.a. skull stripping) is a fundamental step in the neuroimaging pipeline as it can affect the accuracy of downstream preprocess such as image registration, tissue classification, etc. Most brain extraction tools have been mainly orientated for human data and are often challenging for non-human primates (NHP). In recent attempts to improve the performance in NHP, deep learning models appear to outperform the traditional tools. However, given the minimal sample size of most NHP studies and notable variations in data quality, the deep learning models are very rarely applied in multi-site samples in NHP imaging. To overcome this challenge, we propose to use transfer-learning framework that leverages a large human imaging dataset to pretrain a convolutional neural network (i.e. U-Net Model), and then transferred to NHP data using a small NHP training sample. The resulting transfer-learning model converged faster and achieved more accurate performance than a similar U-Net Model trained exclusively on NHP samples. We improved the generalizability of the model by upgrading the transfer-learned model using additional training datasets from multiple research sites in the Primate Data-Exchange (PRIME-DE) consortium. Our final model outperformed brain extraction routines from popular MRI packages (AFNI, FSL, and FreeSurfer) across multiple heterogeneous multiple sites from PRIME-DE with less computational cost (20s~10min). Our model, code, and the skull-stripped mask repository of 136 macaque monkeys are publicly available for unrestricted use by the neuroimaging community at https://github.com/HumanBrainED/NHP-BrainExtraction.

2Compelling evidence suggests the need for more data per individual to reliably map the functional 3 organization of the human connectome. As the notion that 'more data is better' emerges as a golden 4 rule for functional connectomics, researchers find themselves grappling with the challenges of how 5 to obtain the desired amounts of data per participant in a practical manner, particularly for 6 retrospective data aggregation. Increasingly, the aggregation of data across all fMRI scans 7 available for an individual is being viewed as a solution, regardless of scan condition (e.g., rest, 8 task, movie). A number of open questions exist regarding the aggregation process and the impact 9 of different decisions on the reliability of resultant aggregate data. We leveraged the availability 10 of highly sampled test-retest datasets to systematically examine the impact of data aggregation 11 strategies on the reliability of whole-brain functional connectomics. Specifically, we compared 12 functional connectivity estimates derived after concatenating from: 1) multiple scans under the 13 same state, 2) multiple scans under different states (i.e. hybrid or general functional connectivity), 14 and 3) subsets of one long scan. We also varied connectivity processing (i.e. global signal 15 regression, ICA-FIX, and task regression) and estimation procedures. When the total number of 16 time points is equal, and the scan state held constant, concatenating multiple shorter scans had a 17 clear advantage over a single long scan. However, this was not necessarily true when concatenating 18 across different fMRI states (i.e. task conditions), where the reliability from the aggregate data 19 varied across states. Concatenating fewer numbers of states that are more reliable tends to yield 20 higher reliability. Our findings provide an overview of multiple dependencies of data 21 concatenation that should be considered to optimize reliability in analysis of functional 22 connectivity data.

No abstract