External validation of the RSNA 2020 pulmonary embolism detection challenge winning deep learning algorithm

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Pulmonary embolism (PE) diagnosis relies on computed tomography pulmonary angiography (CTPA).
• Deep learning (DL) algorithms show promise for automated PE detection.
• Evaluating DL algorithm generalizability across different CT technologies is crucial.

Purpose of the Study:

• To assess the diagnostic performance of a leading DL algorithm for PE detection.
• To evaluate the algorithm's generalizability to local patient data from two hospitals.
• To compare performance on multidetector CT (MDCT) versus spectral detector CT (SDCT) images.

Main Methods:

• Retrospective analysis of CTPA images from patients with suspected PE.
• Retraining the RSNA 2020 challenge-winning DL algorithm on the RSPECT dataset.
• Testing the algorithm on MDCT (Hospital A) and SDCT/VMI (Hospital B) datasets, comparing against radiologist consensus.

Main Results:

• The DL algorithm achieved high diagnostic accuracy, with an area under the curve (AUC) of 0.96 in Hospital A (MDCT).
• Performance was slightly lower in Hospital B, with AUCs of 0.89 for conventional reconstruction and 0.87 for VMI on SDCT.
• PE prevalence was 30.7% in Hospital A and 29.0% in Hospital B.

Conclusions:

• The retrained DL algorithm demonstrates high diagnostic accuracy for PE detection on MDCT images.
• A slight decrease in performance on SDCT images suggests a need for adaptation to newer CT technologies.
• Further training on diverse CT data may improve the generalizability and robustness of DL algorithms for PE detection.