Accuracy of a deep learning-based algorithm for the detection of thoracic aortic calcifications in chest computed tomography and cardiovascular surgery planning

Comprehensive Summary

Saffar et al. retrospectively evaluated a deep learning algorithm for automated detection of thoracic aortic calcifications (TAC) on chest CT scans in 91 patients (100 scans) undergoing cardiothoracic surgery assessment. The system segmented the aorta into eight regions and quantified calcification volumes, with visual binary ratings as reference. TAC were present in 74% of patients. Optimal parameters for detecting calcifications in the aortic clamping zone achieved a sensitivity of 93%, specificity of 82%, and area under the ROC curve of 0.874. Inter-rater agreement between algorithm and human readers ranged from κ = 0.66 to 0.92 across segments, and complete results were obtained in 92-95% of scans with 5-7 minutes of processing time. False positives were associated with older age and heavy calcification burden, while false negatives occurred more often in mildly affected segments.

Outcomes and Implications

The algorithm showed high accuracy for identifying calcified plaques in the thoracic aorta, including the aortic clamping zone, suggesting potential value for preoperative cardiovascular surgery planning. Automated mapping could standardize TAC evaluation and reduce observer variability, but the single-center retrospective design and limited cohort size restrict generalizability. Missed microcalcifications and segment-boundary errors indicate the need for further technical refinement and multicenter validation before clinical implementation.