In a retrospective paired-cohort study of 716 same-day NCCT/CTPA exams (129 PE-positive, 587 PE-negative), a deep learning model detected pulmonary embolism on noncontrast CT with an AUC of 0.97, accuracy of 0.9306, sensitivity of 0.9655, and specificity of 0.7857. The study suggests NCCT alone is no longer enough for exclusion, but AI on NCCT may be enough for decision support when CTPA is contraindicated.

Why It Matters To Your Practice

• PE is high-stakes and often nonspecific clinically, yet CTPA cannot be performed in every patient because of contrast-related contraindications.

• This study used paired same-day scans from the same patients, helping isolate what AI could learn from NCCT when thrombus location was confirmed on CTPA.

• If validated externally, AI-assisted NCCT could add value in settings where clinicians currently have limited imaging options.

Clinical Implications

• For patients who cannot receive contrast, an AI flag on NCCT could help prioritize further workup, closer monitoring, or expedited alternative imaging.

• The strongest NCCT model, AttnFocus-3D-2, achieved high sensitivity, which may be useful for triage-oriented workflows where missed PE is the main concern.

• Specificity was more modest at 0.7857, so positive AI findings would still need clinical correlation and should not replace confirmatory diagnostic pathways.

Insights

• The dataset included 716 patients scanned between 2018 and 2026, with 70%/20%/10% patient-level splits for training, validation, and independent testing.

• Scans were converted into 12-second mediastinal-window videos, and a 40-frame segment centered at the temporal midpoint was analyzed.

• Among three 3D architectures tested, AttnFocus-3D performed best overall; on CTPA, the top model reached AUC 0.98, while a BiLSTM-ResNet-3D variant had the most balanced sensitivity-specificity profile.

The Bottom Line

• AI may make noncontrast chest CT more clinically useful for PE decision support, especially when CTPA is not feasible.

• Do not treat this as a rule-out tool yet: the results are promising but retrospective and need external, real-world validation before changing standard diagnostic practice.

• Near-term impact is most likely in triage and diagnostic uncertainty, not autonomous PE diagnosis.

Reference: Video-based Deep Learning for Pulmonary Embolism Classification on Paired Contrast-enhanced and Noncontrast Chest CT: A Single-center Retrospective Feasibility Study.