Comparison of Human-in-the-Loop Neural Network and Manual Methods for Aortic Diameter Measurement at CT Angiography

Purpose To compare the performance of a convolutional neural network (U-Net) with human-in-the-loop (HITL) validation against manual clinical measurements of thoracic aortic diameter at CT angiography. Materials and Methods This retrospective analysis included patients with thoracic aortic dilatation and at least two CT angiographic examinations between January 2006 and April 2023. Manual diameters were measured by technicians trained in the three-dimensional method. A multitask U-Net performed aortic segmentation, landmark localization, and automated aortic diameter measurements, followed by HITL validation. Discrepancies in diameter greater than 5 mm underwent expert remeasurement. Mid ascending aortic growth from U-Net and manual measurements were compared against a diameter-independent three-dimensional method (vascular deformation mapping). Agreement was assessed using Bland-Altman analysis and intraclass correlation coefficients. Results This study included 177 patients (101 [57%] male patients; median age, 64 years [IQR, 57-71]). Among 2028 paired measurements, 1955 (96%) passed the HITL validation. Validation reduced the 95% limits of agreement from -4.2 to 4.3 mm to -2.9 to 3.4 mm and reduced large discrepancies by 61% (from 61 to 23 measurements; P < .01). Expert remeasurements of discrepancies showed lower mean difference ± SD compared with U-Net measurements (2.2 mm ± 1.7 vs 7.0 mm ± 2.9; P < .01). U-Net measurements were more likely to yield negative growth values than manual measurements (odds ratio, 0.56; 95% CI: 0.44, 0.70; P < .01). Compared with vascular deformation mapping-derived growth, U-Net measurement showed stronger agreement than manual measurement (intraclass correlation coefficient, 0.74 [95% CI: 0.62, 0.80] vs 0.40 [95% CI: 0.18, 0.56]; P < .01). Conclusion Automated U-Net measurements of thoracic aorta diameters, when validated with an HITL approach, demonstrated stronger agreement with reference standard assessment than manual clinical measurements. Keywords: Aorta, Neural Networks, Vascular, CT Angiography, Segmentation, Thoracic Aortic Diameter Supplemental material is available for this article. © RSNA, 2025.