Image quality is not an issue for AI model that detects pulmonary embolisms on CT
Artificial intelligence models are effective at triaging computed tomography pulmonary angiography (CTPA) scans for pulmonary embolisms (PE), regardless of image quality.
In a study published in Clinical Imaging, a commercially available AI model achieved high sensitivity and specificity for PE detection, even when the diagnostic quality of the CT scans was not perfect, according to the interpreting radiologist.
“CTPA can be suboptimal for PE evaluation due to suboptimal contrast opacification and/or artifacts related to motion, large body habitus, and major orthopedic hardware,” corresponding author Shadi Ebrahimian, from the department of radiology at Massachusetts General Hospital and Harvard Medical School, and co-authors disclosed. “Suboptimal CTPA can lead to repeat scanning, increased radiation doses, increased contrast media requirement and reinjection, and delay in care.”
Pulmonary embolisms have significant morbidity and mortality rates when not diagnosed and treated promptly. CTPA scans are the standard of care for diagnosing PEs, but suboptimal scans make it difficult for radiologists to evaluate images. Some triage algorithms have been shown to detect PE on CTPA scans, but their effectiveness has never been analyzed when the quality of the exam is suboptimal.
This is what led researchers to testing an AI-based, commercially available, FDA-cleared PE triage algorithm’s ability to detect PEs on suboptimal CT scans. For the study, researchers identified 104 CTPA scans that had been deemed suboptimal for PE detection in radiology reports. They added an additional 226 optimal scans both with and without PE presence and had two thoracic radiologists review all exams for adequacy and presence/location of blood clots.
Out of 226 scans considered to be optimal, 47 (21%) were reclassified as suboptimal after the triage algorithm was applied. Of the 330 total exams, PEs were identified on 97. The AI triage system had similar sensitivity and specificity for PE detection on both sets of scans, at 100% and 89% on the suboptimal set and 96% and 92% for the optimal set. The algorithm also detected several clots that were missed on the original reads.
“The AI algorithm for detection of PE retained a high level of performance at par with the radiologists regardless of the technical adequacy of CTPA examinations,” the researchers noted. “The major implication of our study is the demonstration of how an algorithm such as the one used in our study for PE detection can be tested across variable equipment and exam quality. Generalizability across a spectrum of imaging conditions as in our study helps identify strengths and limitations in AI performance.”
The experts suggested that radiologists’ confidence when diagnosing PEs on imaging exams with poor diagnostic quality could benefit from the triage algorithm’s assistive interpretations.
More artificial intelligence content:
AI tool achieves excellent agreement for knee OA severity classification
FDA greenlights AI software that detects fractures and traumatic injuries
Deep learning model triages brain MRIs for abnormalities to prioritize reads
Automated CT scoring system accurately predicts prognosis in stroke patients
Legal ramifications to consider when integrating AI into daily radiology practice
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.