Does an onsite deep learning CT-FFR algorithm accurately detect functionally significant coronary stenosis compared to invasive iFR in patients undergoing CCTA?
An onsite deep learning CT-FFR algorithm provides good diagnostic agreement with invasive iFR and offers incremental value over CCTA alone for assessing coronary stenosis severity.
BACKGROUND Deep learning-based fractional flow reserve derived from coronary CT angiography (CT-FFR) enables noninvasive assessment of lesion-specific ischemia. Onsite CT-FFR systems provide near-real-time physiologic evaluation at the workstation, potentially reducing unnecessary invasive testing. This study evaluated the diagnostic performance of a novel onsite deep learning CT-FFR algorithm compared with invasive instantaneous wave-free ratio (iFR). METHODS We retrospectively analyzed 44 patients (44 lesions) who underwent clinically indicated coronary CT angiography (CCTA) and invasive iFR. CT-FFR values were generated using an onsite deep learning algorithm (cFFR v6) 1-2 cm distal to visually identified stenoses. Physiologic significance was defined as CT-FFR ≤0.80 or iFR ≤0.89. Diagnostic performance metrics were calculated overall and within CCTA stenosis strata (70 %). ROC analysis and Pearson correlation assessed discriminative ability and linear association. Additional comparative analyses evaluated diagnostic accuracy of CCTA ≥50 % and ≥70 % thresholds relative to iFR and quantified incremental diagnostic value of CT-FFR over CCTA alone. RESULTS Of 44 lesions, 28 (63.6 %) were iFR-positive and 30 (68.2 %) were CT-FFR-positive. CT-FFR demonstrated a sensitivity of 89.3 %, specificity of 68.8 %, positive predictive value of 83.3 %, negative predictive value of 78.6 %, and accuracy of 81.8 %; the area under the ROC curve was 0.79 (95 % CI, 0.66-0.92). CT-FFR and iFR showed a modest but significant correlation (r ≈ 0.37). Performance remained favorable in moderate (40-70 %) stenoses (AUC 0.73) and severe (>70 %) stenoses (AUC 0.84). In contrast, CCTA ≥50 % and ≥70 % thresholds showed limited discriminatory ability versus iFR (AUC 0.44 and 0.52, respectively). Compared with CCTA alone, CT-FFR improved both sensitivity and specificity and substantially increased AUC across both thresholds. CONCLUSION The onsite deep learning CT-FFR algorithm demonstrated good diagnostic agreement with invasive iFR and maintained performance across stenosis severity categories, while providing clear incremental value over CCTA stenosis assessment alone. These findings support the feasibility of rapid, workstation-integrated physiologic assessment during CCTA interpretation. Larger multicenter studies are needed to validate these results and clarify the clinical role of onsite CT-FFR.
Roshan et al. (Thu,) studied this question.