Fig 1.
Reliable/unreliable lesion volumes computed at various scan durations.
The arterial input function (AIF) and the vascular output function (VOF) are displayed as reference.
Fig 2.
AIF and VOF derived features used to feed the machine learning algorithms.
AIF: arterial input function; VOF: venous output function; HU: Hounsfield units; UCI: upward contrast increase; DCI: downward contrast increase.
Table 1.
Descriptive statistics of the stable CTP scans.
SD: Scan duration; OSD: Optimal scan duration. Std: standard deviation; P: percentile. AIF: arterial input function. All metrics are reported in seconds.
Fig 3.
Optimal scan duration (OSD) inter-rater analysis.
Left: Histogram of the OSD absolute differences between raters. Right: Bland-Altman plot of the OSD values for the raters. R1: Rater #1. R2: Rater #2.
Fig 4.
Receiver operating characteristic (left) and precision-recall (right) curves.
AUC: area under the curve; SVM_linear: support-vector machine with linear kernel; SVM_rbf: support-vector machine with radial basis function kernel; RF: Random forests; LR: Logistic-regression; Adaboost: Adaptive boosting; Gradboost: Gradient boosting.
Table 2.
Classifiers’ performance for detecting truncation artifacts.
The used operating points are θ = [27, 30, 40, 50, 60] s for the baseline classifier g. Note that 27 s is the optimal operating point for g, defined as the closest point to the ideal classifier with precision = recall = 1. Reported metrics for the machine learning approaches are obtained at the optimal operating points. Outperforming values for each metric are shown in bold. SVMLinear: support-vector machine with linear kernel; SVMRBF: support-vector machine with radial basis function kernel; RF: Random forests; LR: Logistic regression; Gradboost: Gradient boosting classifier. Baseline classifier: threshold on scan duration.
Fig 5.
Relative feature importance for 100 bootstraps with a Gradient boosting classifier.
Bars (error-bars) represent mean (standard deviation). AIF: arterial input function; VOF: venous output function; UCI: upward contrast increase; DCI: downward contrast increase.
Fig 6.
Histograms showing the difference between optimal scan duration and the scan duration for each classifiers’ predicted samples.
Correct predictions comprises scans properly labelled as reliable or unreliable. Incorrect predictions comprises samples wrongly labelled as reliable or unreliable. Inter-rater variability lines are drawn at the 95% inter-rater values (± 2.30 s).
Table 3.
Single features’ classification performance.
The used cutoff is always the optimal operating point, defined as the closest point to the ideal classifier with precision = recall = 1. AIF: arterial input function; VOF: venous output function; UCI: upward contrast increase; DCI: downward contrast increase; ROC: receiver operating characteristic curve; PR: precision-recall curve; AUC: area under the curve.