AI for pulmonary embolism on CT buys speed, not accuracy — three 2026 studies

AI software for pulmonary embolism (PE) scans a CT pulmonary angiogram (CTPA) as soon as it is acquired, flags suspected clots in the lung arteries, and can alert the treating team before a radiologist has opened the study. Three 2026 studies point the same way: the clearest benefit is speed, not accuracy — where detection was measured head-to-head, the AI still missed more clots than radiology trainees, particularly the small peripheral ones [s1][s2][s3].

How the tools work

CTPA is the standard imaging test for suspected PE. Commercial AI triage systems — Viz.ai PE is an FDA-cleared example — run automated detection on the images and generate an alert, with the aim of moving a positive scan to the front of the queue rather than replacing the radiologist's read [s3]. Whether that alerting also improves detection is a separate question from whether it saves time, and the two are best judged apart.

Against radiology residents

The most direct accuracy test analysed all 594 emergency CTPAs done over three months at a university hospital, comparing a commercial AI to on-call radiology residents, with the attending radiologist's final report as the reference standard [s1]. PE was present in 82 patients (13.8% prevalence), of which 41 (50%) were proximally located [s1].

Overall, the AI reached a sensitivity of 89% and specificity of 99%, against 97.6% sensitivity and 99.2% specificity for the residents (differences not statistically significant) [s1]. The split by clot location is where the story is. For proximal PE — the large, central, most dangerous clots — sensitivity was 97.6% for AI and 100% for residents (p = 1) [s1]. For peripheral PE, AI sensitivity fell to 80.5% versus 95.1% for residents (p = 0.08) [s1]. The authors conclude residents performed better overall, with the AI strongest on proximal clots [s1]. A tool that reliably catches central PE but misses one in five peripheral clots is useful as a safety net, not a substitute reader.

Why the numbers may not travel

A second study tested how well a general-purpose model transfers between institutions — the failure mode that has undone many imaging algorithms. It paired the frozen embeddings of Google's CT Foundation model with lightweight classifiers for central PE detection, training on the public RSNA CTPA dataset and externally validating on Stanford's INSPECT cohort [s2]. The best test AUC was 0.79 on the held-out RSNA set [s2]. On the external INSPECT cohort, AUC dropped by 0.17 and specificity at the transferred operating point collapsed, and simple recalibration of the threshold did not restore usable performance [s2]. The authors' conclusion is blunt: frozen generalist embeddings alone do not guarantee cross-institutional reliability [s2]. An accuracy figure from one hospital's data is not a promise about another's.

The speed benefit

The case for AI PE tools rests less on accuracy than on time. A single-institution study reviewed 148 patients who had a PE and underwent mechanical thrombectomy or thrombolysis between July 2018 and March 2025, split into a pre-AI cohort (42 patients) and a post-AI cohort (82 patients) around the hospital's adoption of Viz.ai PE [s3]. Median time to diagnosis — from scan completion to the AI alert, versus to the final radiology report before AI — improved from 72.5 to 40 minutes (P = 0.00005), and median time to intervention fell from 1360 to 1224 minutes (P = 0.036) [s3]. Time to anticoagulation, at 76 versus 61.5 minutes, was not significantly different (P = 0.824) [s3]. Two in-hospital deaths occurred, both in the pre-AI group [s3].

Those are real gains, but the design is a before-and-after comparison across seven years, so improvements in the wider PE pathway over that period — not the AI alone — could account for part of the change. The study measures workflow speed convincingly and mortality only in passing.

What it means, and what to watch

Read together, the three studies describe a tool whose value is triage: it can shave time off the path from scan to treatment for central clots, while trailing human readers on the peripheral emboli that are easiest to miss and whose numbers may not hold in a new hospital. That is a familiar shape — AI imaging devices cleared and adopted on speed and accuracy metrics, with patient-outcome evidence still thin. What to watch is whether faster diagnosis actually changes outcomes in a controlled comparison, and whether detection holds up on external validation rather than a vendor's home data.

For the wider debate, see our coverage of the FDA's AI radiology testing gap, two 2026 radiology-assist studies, AI stroke triage and thrombectomy rates, and what happens after diagnosis in serious PE.

Sources

• Artificial intelligence for pulmonary embolism detection: Is it comparable to radiology residents? — European Journal of Radiology Open, 2026-07-13
• Evaluating the Google CT Foundation model for central pulmonary embolism detection on computed tomography pulmonary angiograms — European Journal of Radiology, 2026-07-20
• Adding artificial intelligence to the pulmonary embolism response team improves time to diagnosis and treatment of pulmonary embolism — Journal of Vascular Surgery: Venous and Lymphatic Disorders, 2026-08-09