2312. Performance of an Artificial Intelligence-Based Platform Against Clinical Radiology Reports for the Evaluation of Non-contrast Chest CT
Authors * Denotes Presenting Author
  1. Basel Yacoub *; Medical University of South Carolina
  2. Ismail Kabakus; Medical University of South Carolina
  3. Joseph Schoepf; Medical University of South Carolina
  4. Vincent Giovagnoli; Medical University of South Carolina
  5. Andreas Fischer; Medical University of South Carolina; University Hospital Basel
  6. Akos Varga-Szemes; Medical University of South Carolina
  7. Tilman Emrich; Medical University of South Carolina; University Medical Center Mainz
To assess the performance of an artificial intelligence (AI) platform against clinical radiology reports on non-contrast chest computed tomography (CT) scans.

Materials and Methods:
Consecutive patients (n=100, 57 men, 64.2 ± 14.8 years) who had previously undergone non-contrast chest CT were retrospectively identified. The original radiology reports were reviewed in a binary fashion for mentioning any of the following findings: pulmonary lesions, pulmonary emphysema, aortic dilatation, coronary artery calcification (CAC), and vertebral compression fracture (VCF). CT scans were processed using a prototype AI platform and the detection of the same findings was recorded. The radiology reports’ findings and the AI results were subsequently compared to a consensus read by two board-certificated radiologists as reference.

Complete AI evaluation of all 5 conditions was possible in 95 CT scans. Aortic segmentation and coronary calcium quantification modules failed to be processed by AI in 2 and 3 cases respectively. AI showed superior diagnostic performance compared to radiology reports in identifying aortic dilatation (sensitivity (SN): 96.3%, specificity (SP): 81.4%, AUC: 0.89) vs (SN: 25.9% SP: 100%, AUC:0.63) p-value <0.001 on DeLong’s test; and on CAC (SN: 89.8%, SP: 100, AUC: 0.95) vs (SN: 75.4%, SP: 94.9%, AUC: 0.85) p-value = 0.005. Reports had better performance than AI in pulmonary lesions (SN: 97.6%, SP: 100%, AUC: 0.99) vs (SN: 92.8%, SP: 82.4%, AUC: 0.88) p-value = 0.024; and on VCF (SN: 100%, SP: 100%, AUC: 1.0) vs (SN: 100%, SP: 63.7%, AUC:0 0.82) p-value <0.001. Similar diagnostic performance was noted in identifying pulmonary emphysema on AI (SN: 80.6%, SP: 66.7%. AUC: 0.74) and reports (SN: 74.2%, SP: 97.1%, AUC: 0.86) p-value = 0.064.

A prototype AI platform had a higher diagnostic performance than radiology reports in identifying aortic dilatation and coronary artery calcifications and showed higher sensitivity for detection of pulmonary emphysema. The implementation of AI into clinical practice AI has promising potential to improve radiology reporting and increase the level of detail in description of the findings present on imaging.