Keywords:
Lung, Pulmonary vessels, Artificial Intelligence, CT-Angiography, CAD, CT, Computer Applications-Detection, diagnosis, Contrast agent-intravenous
Authors:
N. Hendriks1, J. J. Zigterman2, M. Roelofs1, J. Nijboer-Oosterveld1, E. De Boer1, M. F. Boomsma1; 1Zwolle/NL, 2Zwolle, Overijsel/NL
DOI:
10.26044/ecr2019/C-2541
Methods and materials
We included CTPAs from two months scanned on a 256-slice CT-scanner with available raw data and a minimum of 200 Hounsfield Units (HU) in the main pulmonary trunk.
Primary diagnosis was made by the on-call radiologist and for this research the scans were re-evaluated by a thorax-radiologist.
Differences were resolved in a consensus meeting with a third thorax-radiologist and the final diagnosis used as reference standard.
All scans were reconstructed with both HIR (level 4 out of 6 levels) and MBIR (level 1 out of 3 levels) with a slice thickness of 1.0x0.5mm.
Possible pulmonary embolisms were indicated by a region of interest (ROI) by the CAD-software.
All CAD detections were independently reviewed by two radiologists.
Differences were resolved in a consensus meeting.
The CAD detections were categorized as true positive (TP) or false positive (FP).
On every scan the HU and standard deviation were measured with a 90-110 mm2 ROI in the main pulmonary artery and the musculus latissimus dorsi as shown in figure 2.
Interobserver agreement was tested by the Cohen’s Kappa.
Software performance was determined with 2x2 tables describing the sensitivity,
specificity,
positive-predictive value (PPV) and negative predictive value (NPV).
SNR and CNR were calculated according to the following formulas:
SNR = HUvessel / σvessel
CNR = (HUvessel – HUmuscle) / ((σvessel + σmuscle) /2)