An artificial intelligence algorithm for pulmonary embolism detection on polychromatic computed tomography:...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Virtual monochromatic images (VMI) enhance contrast-to-noise ratio in CT scans.
• AI performance may be affected by the shift from conventional polychromatic CT images (CPI) to VMI.
• Established AI algorithms require validation on new imaging techniques.

Purpose of the Study:

• To evaluate the diagnostic performance of an AI algorithm for pulmonary embolism (PE) detection.
• To compare the AI algorithm's performance on conventional polychromatic images (CPI) versus virtual monochromatic images (VMI).
• To assess the reliability of AI trained on older CT data for use with newer VMI technology.

Main Methods:

• Retrospective analysis of 114 patients with suspected PE using paired VMI and CPI from spectral CT.
• An established AI algorithm was used to classify CT pulmonary angiography (CTPA) scans as positive or negative for PE.
• Diagnostic accuracy metrics (sensitivity, specificity, predictive values, likelihood ratios) were compared between VMI and CPI, with a comprehensive reference standard.

Main Results:

• The AI algorithm demonstrated comparable diagnostic accuracy on both CPI and VMI.
• Sensitivity for PE detection was 77.5% on CPI and 85.0% on VMI (p=0.08).
• Specificity was 96.0% on CPI and 94.6% on VMI (p=0.32), with no significant differences observed.

Conclusions:

• The diagnostic performance of the AI algorithm trained on CPI did not significantly decrease when applied to VMI.
• This suggests that commercially available AI algorithms can be safely used on VMI, supporting AI sustainability in CT imaging.
• Ongoing monitoring in clinical practice is recommended despite the reassuring findings.