PERADS.net: Automated PE-RADS Grading with Named Anatomic Localization and Right-to-Left Ventricular Ratio Measurement on CT Pulmonary Angiography

Purpose

To develop an automated pipeline ( PERADS.net ) that segments acute pulmonary embolism on CT pulmonary angiography, assigns a Pulmonary Embolism Reporting and Data System (PE-RADS) grade with named anatomic localization, and measures the right-to-left ventricular (RV/LV) diameter ratio, and to assess radiologist agreement.

Materials and Methods

In this retrospective study, 120 CT pulmonary angiograms (70 peripheral, 35 central, 15 negative) from the public RSNA Pulmonary Embolism CT dataset were analyzed. A two-channel five-fold nnU-Net ensemble segmented the embolus; the pulmonary arterial tree was reduced to a branching graph, and the grade was set by the most proximal level with at least 1% of embolic volume.

Three radiologists, blinded to the algorithm-assigned grade, independently assigned grades and rated RV/LV plausibility. Agreement was assessed with percent agreement and Cohen or Fleiss kappa on five-grade and grouped scales (grade 0 versus 1-2 versus 3-4). The automated ratio was compared with the dataset’s binary RV/LV label in 2119 examinations.

Results

Agreement between the algorithm and three-radiologist consensus (n = 119) was 60.5% (kappa, 0.38) for individual grades and 90.8% (kappa, 0.74; 95% CI: 0.59, 0.87) for the grouped scale; 34 of 47 discordant examinations fell within grades 3-4. As a fourth reader, the algorithm matched grouped-scale interobserver agreement (mean kappa, 0.62 versus 0.67). Radiologists showed no agreement rating RV/LV plausibility as favorable or incorrect (Fleiss kappa, −0.00), whereas the automated ratio agreed with the external label (kappa, 0.49), overestimating strain 3.1:1.

Conclusion

Automated PE-RADS grading agreed with radiologist consensus comparably to a fourth reader on the grouped scale.

Summary Statement

An automated pipeline assigned PE-RADS grades agreeing with three-radiologist consensus at a level approaching interobserver agreement on the clinically grouped scale, while automated RV/LV measurement showed no reader agreement on plausibility.

Key Points

• In 120 CT pulmonary angiograms, agreement between automated and consensus PE-RADS grading was 90.8% (kappa, 0.74; 95% CI: 0.59, 0.87) on the clinically grouped scale (grade 0 versus 1-2 versus 3-4) and 60.5% (kappa, 0.38; 95% CI: 0.24, 0.51) on the five-grade scale.

• Treated as a fourth reader, the algorithm reached grouped-scale agreement with individual radiologists (mean kappa, 0.62) comparable to that observed among radiologists themselves (mean kappa, 0.67).

• Automated right-to-left ventricular ratio agreed moderately with an independent binary label in 2119 examinations (kappa, 0.49; 95% CI: 0.45, 0.52), whereas three radiologists showed no agreement when rating the same measurement as favorable or incorrect (Fleiss kappa, −0.00; 95% CI: −0.09, 0.09).