Improved computer-aided detection performance for pulmonary embolism on CTPA with higher noise reduction reconstruction

To investigate whether a higher image noise reduction improves the diagnostic accuracy of Computer-Aided Detection (CAD) software to detect pulmonary embolism (PE) on CT pulmonary angiography (CTPA).

We retrospectively included 238 consecutive CTPAs of patients with suspected PE obtained between 01/09/2014 and 06/03/2014. Hybrid iterative reconstruction (HIR) was performed using either iDose⁴ level 3 or level 4 with 23% vs. 29% noise reduction, respectively. CAD software marked all potential PE. Two radiologists evaluated CAD markers and classified them as either true positive (TP) or false positive (FP). The reference standard was determined by a consensus reading of two experienced radiologists.

In total, 110 scans made use of iDose⁴ level 3 HIR and 128 scans made use of iDose⁴ level 4 HIR. PE was present in 34 patients in the iDose⁴ level 3 group (30.9%) and in 39 patients in the iDose⁴ level 4 group (30.5%). Sensitivity of CAD software was not significantly different between noise reduction groups level 3 and level 4 (100% and 92.3% respectively, p = 0.24). Specificity was significantly higher in the level 4 group compared to the level 3 group (48.3% vs. 11.8%, p < 0.001). CNR was significantly higher in CT images with level 4 noise reduction (12.0 vs. 9.8, p < 0.001).

CAD software produced significantly fewer FP PE markers in images with higher CNR, reconstructed using a higher noise-reduction setting. The noise reduction level 4 images had significantly higher CNR values, indicating that image quality has a substantial impact on the performance of CAD software.