Fig 1.
Example images with different image quality (IQ) scores.
Poor quality (IQ = 0) due to incorrect slice location (A) and extreme imaging artefacts (B). 0 < IQ < 1 due to incorrect slice location (C) and severe artefacts (D). Major issues (IQ = 1) due to slice location (E) and artefacts (F). Example images with (G) minor issues (IQ = 2) and (H) good quality (IQ = 3). Please refer to the methods section in the main text for their definitions.
Fig 2.
Automated Ascending and Proximal Descending Aorta (AA and PDA) detection-localization and quality control.
Fig 3.
Graphical representation and list of local features (A-R) extracted for each candidate ROI (yellow circles) and divided into groups. Spatial features: diagonal axes A-B and distance mask C of the body. Shape features: J is indicated by the line inside the circles; K is represented by the red circles on the Motion Periodicity (MP) map; L and M describe the circles detected by the Circular Hough Transform (CHT); and N-R characterize the Maximally Stable Extremal Regions (MSER) marked by the red contours. Graphical representation was not possible for features D-I. Please refer to the methods section in the main text for the definition of each image feature.
Fig 4.
Optimization of the random forest (RF) parameters.
Out-of-bag (OOB) mean classification error and CI were calculated from RF training repetitions. A) Training set size for RF with 50 and 1000 trees. B) Number of features per decision split for RF with 50 and 1000 trees. C) Number of trees for RF with 6 features per split. Total number of scans in the training dataset = 1200.
Fig 5.
ROIs classified as Ascending Aorta (AA), Proximal Descending Aorta (PDA) or Not Aorta (NA). Total number of scans in the training dataset = 1200.
Fig 6.
Absolute correlation between pairs of image features (A) and relative importance (B). Calculated from out-of-bag (OOB) observations in the training dataset (1200 scans in total).
Table 1.
Confusion matrix for predicted and actual ROI classes.
Mean classification error = 0.44% (CI = 0.41–0.45%) estimated from RF training repetitions on the test dataset (51482 ROIs).
Fig 7.
A) Receiver Operating Characteristic (ROC) and C) Precision-Recall (PR) analysis for AA and PDA detection-localization. Mean curves and 95% CIs were obtained from bootstrap replicas. Precision baseline was 98.5% for AA and 99.6% for PDA. B) Dice Similarity Coefficient (DSC) for AA and PDA True Positives was ≥ 0.9 in 94.8% and 99.5% of cases, respectively. Total number of scans in the test dataset = 3900.
Fig 8.
Confusion matrix for AA detection (with example images).
Mean error = 0.64% (CI = 0.54–0.74%) estimated from RF training repetitions. Green circles represent the ground truth and red circles the automatically detected ROIs. Total number of scans in the test dataset = 3900.
Fig 9.
Confusion matrix for PDA detection (with example images).
Mean error = 0.18% (CI = 0.15–0.26%) estimated from RF training repetitions. Green circles represent the ground truth and red circles the automatically detected ROIs. Total number of scans in the test dataset = 3900.
Fig 10.
Automated AA×PDA detection probability grouped by CINE image quality (IQ).
Total number of scans in the test dataset = 3900. The difference among the highest IQ groups and between the 2 lowest IQ groups is not significant (ns).
Fig 11.
A) ROC and B) PR analysis for quality control. Mean curves and CIs (from bootstrap replicas) for 2 possible data partitions (using cut-off at IQ ≥ 1 or at IQ > 1). Precision baseline was 96% for IQ ≥ 1 and 87% for IQ > 1.