Bias analysis: Can AI tool handle aortic segmentation and measurements in patients with Scoliosis?

Purpose: Unreported or missed aortic aneurysm can lead to substantial morbidity, mortality, and medicolegal implications. We stress tested an AI tool for aortic segmentation and measurements on contrast-enhanced CT angiography (CTA) to determine performance bias in patients with scoliosis.

Methods/Materials: Our retrospective IRB approved study included 246 CTA from adult patients (>21 years; F:M 158:88) with scoliosis based on Cobb angle measurement (mild < 20º, 49 patients; moderate 20-40º, 144 patients; severe scoliosis>40º, 53 patients) from 10 US sites. We enriched with 400 CTA from adult patients without scoliosis. Based on the manual 3-plane aortic measurements, we classified dataset into normal aorta (NA), aortic ectasia (AE), and aortic aneurysm (AA). All de-identified thin-section CT images (1-1.25mm) were processed with the AI tool (Siemens AI Rad Companion) and the outputs were recorded as DICOM captures and CSV file format. Data were analyzed to estimate sensitivity, specificity, and area under the curve (AUC) for receiver operating characteristic (ROC) analysis.

Results: The respective AUC, sensitivity, and specificity for ascending AA/AE in patients with mild scoliosis were 0.91, 0.94, and 0.80, and 0.80, 0.68, and 0.77 in severe scoliosis (p < 0.01). The respective AUC, sensitivity, and specificity for descending AA/AE in patients with mild scoliosis were 0.84, 0.54, and 0.78, whereas in patients with severe scoliosis were 0.67, 0.97, and 0.3. The corresponding statistics for patients without scoliosis were 0.86-0.88, 0.74-0.96, and 0.81-0.97, for AE and AA, respectively.

Conclusions: There is a small but significant drop in performance of the AI tool in aortic measurements in patients with severe scoliosis as compared to those with mild-moderate scoliosis and normal spinal curvature.
Clinical relevance: Radiologists must be aware of underlying biases in the performance of the AI tool in patients with anatomic variations.