Objectives: To evaluate if dual-energy CT (DECT) pulmonary angiography (CTPA) can detect anemia with the aid of machine learning.
Methods: Inclusion of 100 patients (mean age ± SD, 51.3 ± 14.8 years; male-to-female ratio, 42/58) who underwent DECT CTPA and hemoglobin (Hb) analysis within 24 h, including 50 cases with Hb below and 50 controls with Hb ≥ 12 g/dL. Blood pool attenuation was assessed on virtual noncontrast (VNC) images at eight locations. A classification model using extreme gradient-boosted trees was developed on a training set (n = 76) for differentiating cases from controls. The best model was evaluated in a separate test set (n = 24).
Results: Blood pool attenuation was significantly lower in cases than controls (p-values < 0.01), except in the right atrium (p = 0.06). The machine learning model had sensitivity, specificity, and accuracy of 83%, 92%, and 88%, respectively. Measurements at the descending aorta had the highest relative importance among all features; a threshold of 43 HU yielded sensitivity, specificity, and accuracy of 68%, 76%, and 72%, respectively.
Conclusion: VNC imaging and machine learning shows good diagnostic performance for detecting anemia on DECT CTPA.
Keywords: anemia; computed tomography angiography; machine learning; pulmonary embolism.