Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Correlation plot between MRIQC IQMS (columns) and CAT12 quality measures (rows). MRIQC generated total 68 IQMs and from CAT12 we extracted 36 quality measures.

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques: Generated

Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Balanced accuracy of proposed classifiers, MRIQC and CAT12 on combined and site-wise test data. The model’s performance on the test data is assessed across different (ranked) feature sizes (refer ), with the displayed plot here representing only the best performance selected across these feature sizes. Number of samples in the test data are provided in brackets for each dataset (x-axis). Note that three sites (ADNI GE 3T, ADNI Philips 1.5T, and ADNI Siemens 2.9T) are not included in the figure due to the absence of samples in the reject class resulting in NaN values for balanced accuracies.

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques:

Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Balanced accuracy of proposed classifiers, MRIQC, and CAT12 analysed separately for scans from healthy individuals, patients, and each diagnostic sub-category within both healthy and patient groups in the test dataset. The model’s performance on the test data is assessed across different (ranked) feature sizes, with the displayed plot representing only the best performance selected across these feature sizes (refer ). Number of samples in the test data are provided in brackets for each category (x-axis). Legend of diagnostic subgroups: CN = cognitively normal; HC = healthy controls; MCI = mild cognitive impairment; RBD = REM sleep behaviour disorder; iPD = idiopathic Parkinson’s disease.

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques: Diagnostic Assay

Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Scatter plots for two features snr-total from MRIQC on x-axis and noiseNCR-rps from CAT12 on y-axis showing different levels of overlap for different combinations of dataset, field strength and manufacturer: a) same dataset (Whitehall2), manufacturer (Siemens) and field strength (3T) but different scanner models; b) same scanner model (Siemens 3T Prisma) but different datasets; c) same manufacturer and field strength (Siemens 3T) but different datasets; d) same dataset (ADNI) and field strength (3T) but different manufacturers; e) same dataset (ADNI) and manufacturer (Siemens) but different field strengths; f) different datasets, manufacturers and field strength.

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques:

Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Balanced accuracy of MRIQC, CAT12, and the proposed RUS classifier for leave-one-site-out models. The model’s performance on the test data is assessed across different (ranked) feature sizes (refer in the Supplementary Document), with the displayed plot here representing only the best performance selected across these feature sizes. The total number of samples for each test site is provided in brackets (x-axis). For RUS classifier, each site was kept as test data and classifier was trained on remaining sites using the hyperparameters and feature ranking derived from combined data model (best model with 80 feature size). For reference, we also provide the balanced accuracy of RUS classifier for each site within the test data of the combined data model to see how well our classifier generalises to test data from different sites (diamond marker with grey). Note that balanced accuracies for the combined data model are not included for three sites (ADNI GE 3T, ADNI Philips 1.5T, and ADNI Siemens 2.9T) due to the absence of samples in the reject class of the test data (resulting in NaN values for balanced accuracies).

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques: Derivative Assay, Marker

Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Balanced accuracy of MRIQC, CAT12, and the proposed RUS classifier for exploratory models on field strengths and manufacturers. The model’s performance on the test data is assessed across different (ranked) feature sizes (refer in the Supplementary Document), with the displayed plot here representing only the best performance selected across these feature sizes. The total number of samples for each test site is provided in brackets (x-axis). Field strength: performance of models trained on 3T scanners data (Siemens, Philips, GE) and tested on 1.5T (Siemens, Philips, GE) and 2.9T (Siemens) scanners data; manufacturer: performance of models trained on Siemens (1.5T, 2.9T, 3T) data and tested on Philips (1.5T, 3T) and GE (1.5T, 3T) data; manufacturer and field strength: performance of models trained on Siemens 3T data and tested on Siemens (1.5T, 2.9T), Philips (1.5T, 3T), and GE (1.5T, 3T) data. Additionally, the balanced accuracy of the RUS classifier within the test data for the combined data model for each scenario is presented for reference (diamond marker with grey).

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques: Marker

Journal: Imaging Neuroscience

Article Title: Automated quality control of T1-weighted brain MRI scans for clinical research datasets: methods comparison and design of a quality prediction classifier

doi: 10.1162/IMAG.a.4

Figure Lengend Snippet: Balanced accuracy of MRIQC, CAT12, and the proposed RUS classifier for exploratory models on the MR-ART dataset. We tested the option of applying our combined data model without adding MR-ART samples (RUS—without MRART training) and after retraining including a proportion of MR-ART images. X-axis indicates the % training samples used in addition to main training data ( N = 1955) and total number of training and test samples from MR-ART (details in ). Feature ranking, feature size, and hyperparameters were taken from the combined data model.

Article Snippet: Each defaced T1w brain image was processed in CAT12 segmentation pipeline (standalone version r2042 running on v93 of MATLAB compiler runtime).

Techniques: