Quantification of the Thoracic Aorta and Detection of Aneurysm at CT: Development and Validation of a Fully Automatic Methodology

Abstract

Purpose

To develop and validate a deep learning–based system that predicts the largest ascending and descending aortic diameters at chest CT through automatic thoracic aortic segmentation and identifies aneurysms in each segment.

Materials and Methods

In this retrospective study conducted from July 2019 to February 2021, a U-Net and a postprocessing algorithm for thoracic aortic segmentation and measurement were developed by using a dataset (dataset A) that included 315 CT studies split into training, hyperparameter-tuning, and testing sets. The U-Net and postprocessing algorithm were associated with a Digital Imaging and Communications in Medicine series filter and visualization interface and were further validated by using a dataset (dataset B) that included 1400 routine CT studies. In dataset B, system-predicted measurements were compared with annotations made by two independent readers as well as radiology reports to evaluate system performance.

Results

In dataset B, the mean absolute error between the automatic and reader-measured diameters was equal to or less than 0.27 cm for both the ascending aorta and the descending aorta. The intraclass correlation coefficients (ICCs) were greater than 0.80 for the ascending aorta and equal to or greater than 0.70 for the descending aorta, and the ICCs between readers were 0.91 (95% CI: 0.90, 0.92) and 0.82 (95% CI: 0.80, 0.84), respectively. Aneurysm detection accuracy was 88% (95% CI: 86, 90) and 81% (95% CI: 79, 83) compared with reader 1 and 90% (95% CI: 88, 91) and 82% (95% CI: 80, 84) compared with reader 2 for the ascending aorta and descending aorta, respectively.

Conclusion

Thoracic aortic aneurysms were accurately predicted at CT by using deep learning.

Keywords: Aorta, Convolutional Neural Network, Machine Learning, CT, Thorax, Aneurysms

Supplemental material is available for this article.

© RSNA, 2022

Summary

A deep learning–based system for automatic thoracic aortic measurement was developed to prevent aneurysm underreporting in chest CT performed for noncardiovascular complaints and was validated in a large chest CT dataset.

Key Points

• ■ A U-Net and a postprocessing algorithm were trained and tested on a dataset consisting of 315 chest and chest plus abdomen CT studies (with and without contrast material enhancement) and CT angiographic studies.

• ■ After being linked to other components (a rule-based series filter and a visualization interface), the whole system was validated on another dataset consisting of 1400 CT studies, which were routine studies of the chest or studies of the chest plus abdomen with and without contrast enhancement.

• ■ When tested on an independent dataset acquired from a different time period than that used for the main dataset, the system achieved accuracy of 88% (95% CI: 86, 90) and 90% (95% CI: 88, 91) for aneurysm detection in the ascending aorta and achieved accuracy of 81% (95% CI: 79, 83) and 82% (95% CI: 80, 84) for aneurysm detection in the descending aorta (when using the annotations from readers 1 and 2 as references, respectively).

• ■ The system achieved a mean absolute error between the automatic and reader-measured diameters equal to or less than 0.27 cm for both the ascending aorta and the descending aorta, thus showing the generalization capacity and potential use for thoracic aneurysm triaging and measuring purposes.

Introduction

Thoracic aortic aneurysm (TAA) refers to enlargement of the aorta by more than 50% of the normal diameter in segments between the aortic valve and aortic hiatus of the diaphragm (1). CT and MR angiography are the optimal imaging modalities to detect TAAs, determine the aortic diameter, and identify rupture or dissection (2). Recent data suggest that the prevalence of TAAs varies between 0.16% and 0.34% (3), with approximately 60% of TAAs involving the ascending aorta, 40% involving the descending aorta, 10% involving the arch, and 10% involving the thoracoabdominal aorta (4,5). However, because of the inherently silent nature of TAAs, with greater than 95% of patients being asymptomatic until the aneurysm ruptures (6), this incidence is likely underestimated, and most TAAs remain undetected unless incidentally discovered (1,5).

The prevalence of incidental thoracic aortic dilatation ranges from 2.2% in nonangiographic (routine) chest CT in the general population (7) to 23% in electrocardiographically gated chest CT in patients with atrial fibrillation (8). As CT imaging becomes more available and lung cancer screening becomes more frequent, radiologists encounter more incidental findings (9). Yet it remains unclear whether radiologists evaluate and report cardiovascular abnormalities from nonvascular chest CT, with disclosed rates being as high as 63% for unreported cardiovascular findings from chest CT performed for pulmonary evaluation (10). Furthermore, because of the potential of TAAs to cause severe complications, such as dissection and rupture, that are associated with high mortality rates (between 94% and 100% for aortic rupture) (11), any chest imaging should ideally be leveraged for investigation of an aneurysm, especially if the finding is not previously known.

In this work, we aimed to develop and validate a fully automatic artificial intelligence (AI)–based system to predict the diameters of the ascending aorta and descending aorta with good agreement with the diameters measured by human readers. The system, working entirely unsupervised, was used to determine the best image series to analyze within a complete CT study, segment the regions of the aorta in that series, extract multiple diameter measurements along the entire length of the ascending and descending thoracic aorta, and output the largest diameter relative to the centerline of the aorta for each aortic segment. The predicted diameters were then used to classify the study according to the presence or absence of aneurysms on the basis of predetermined thresholds.

Materials and Methods

Cohort Selection

This was a Health Insurance Portability and Accountability Act–compliant study that was conducted from July 2019 to February 2021 and was approved by the institutional review board with a waiver of patient consent. The cohort consisted of two datasets of chest and chest plus abdomen CT (routine) and CT angiography studies, which were acquired retrospectively from the picture archiving and communication system of a single institution.

Dataset A (the main dataset) was gathered for model development and testing with the postprocessing algorithm. The data included routine CT, with or without contrast material, of the chest or chest plus abdomen and CT angiography of the chest or chest plus abdomen performed from May 2019 to June 2019.

Dataset B (the validation dataset) contained consecutive routine CT studies with and without contrast material that were performed between 2007 and 2018 and was collected after model development. This dataset was used only for testing the whole system. Routine chest CT studies focus on the study of nonvascular structures. Compared with chest CT angiographic studies, they have lower resolution, and contrast-enhanced series are acquired under a less strict time frame following contrast material injection. As a result, they are more aligned with our intended use, which is to alert the radiologist to the presence of an aneurysm at chest CT performed for nonvascular clinical reasons, favoring incidental aneurysm detection.

Both datasets were acquired with General Electric, Philips, and Siemens scanners. No distinction was made between aneurysms with and without dissection, and no studies were excluded based on image quality, so as to reflect the heterogeneity of studies observed in routine clinical practice. Data were de-identified by using the gdcmanon tool (http://gdcm.sourceforge.net/html/gdcmanon.html).

Dataset A

For the main dataset, 320 CT angiographic and routine CT studies were selected. Two duplicate studies from the same patient were excluded, and three studies were excluded because of inadequate patient position (ie, the patient was not in the supine position). The resulting dataset consisted of 315 studies; 145 were used for training the segmentation model, 15 were used for hyperparameter tuning, and 155 were used for testing. The characterization of this sample is provided in Table 1. The testing subset of this sample comprised 40 Digital Imaging and Communications in Medicine (DICOM) series from patients with an aneurysm in the ascending aorta, 29 DICOM series from patients with an aneurysm in the descending aorta, 40 DICOM series from patients with an aneurysm in both the ascending and descending aortic regions, and 46 DICOM series from patients without an aneurysm.

Table 1:

Characteristics of Dataset A

Dataset B

For evaluation of the entire system, 1632 routine CT studies were selected, and the following studies were excluded from the cohort: 168 were duplicate scans from repeat patients, 30 had missing images, 19 were incomplete or inadequate studies (eg, only the lung window or maximum intensity projection), 10 were from patients younger than 18 years of age, two had inadequate patient positioning, and three showed saccular aneurysms (thus, the largest aortic diameter did not correspond to the aneurysm diameter). The final validation set comprised 1400 studies from separate patients. In this dataset, the number of DICOM series (ie, a sequence of images consecutively acquired) in each study ranged from one to 30, with a median of seven series being included. The characterization of this sample is provided in Table 2.

Table 2:

Characteristics of Dataset B

Dataset Annotation

The presence or absence of aortic aneurysm(s), their location(s), and available measurements were obtained from the radiology reports for both datasets. For dataset A, a radiologist (F.B.C.M.) with more than 7 years of experience in reading chest CT studies selected a single DICOM series from each CT study to train and test the AI model along with the postprocessing algorithm. The series selection was similar to the selection that would be performed by a radiologist for the task of aortic aneurysm detection, and it prioritized the axial series with the highest resolution and with contrast enhancement in the following manner: (a) if a contrast-enhanced series was not available, the series with the largest number of images was selected; (b) if a contrast-enhanced series was available, the contrast-enhanced series was selected; and (c) if multiple series with contrast enhancement were available, then the contrast-enhanced series in the arterial phase was selected. The same DICOM series selection process was performed in dataset B by two independent readers: a research fellow (A.T.) with 2 years of experience in cardiovascular imaging research (reader 1) and the previously mentioned radiologist (reader 2). In dataset B, the selected series was used to measure the largest ascending and descending aortic diameters and also served as a benchmark for evaluating the series selection filter.

For dataset A, fully manual contouring of the aorta was performed on all DICOM series in the training set (145 series) and on 50 series of the test set. Four radiologists, each with more than 5 years of experience, drew multiple connected points delineating (and including) the vessel’s outer wall for each image in the series, from the aortic root (including) up to one image above the origin of the celiac trunk. The annotations were made by using an internally developed annotation tool and were reviewed by a fifth radiologist (F.B.C.M.) to ensure accurate ground truth contours (Fig 1).

Figure 1:

Axial contrast-enhanced CT images from a study that was used to train the U-Net. The yellow circles indicate the manual contouring of the aorta’s outer wall (A) at the top of the arch, (B) at the level of the pulmonary veins, and (C) one level above the celiac trunk’s origin.

Other types of annotations were also performed for dataset B. First, the largest diameter of the ascending and descending aortas was independently measured from the images of the selected DICOM series by readers 1 and 2. Both readers followed a consistent annotation procedure: (a) multiplanar reformation, (b) visual inspection of the whole vessel, (c) identification of the largest portion in each segment, and (d) measurement of the maximal diameter in the plane perpendicular to the centerline of the aorta at these levels. In the case of multiple aneurysms, only the largest aneurysm in each aortic segment was measured. Second, to benchmark the automatic division of the aorta performed by the postprocessing algorithm, a random sample of 100 DICOM series was selected to obtain the x, y, and z coordinates of the point of transition between the ascending aorta and the descending aorta. This point was defined as the point at which the line parallel to the lateral margin of the left subclavian artery crossed the center of the aorta. In both datasets, the ascending aorta was defined as the segment from the aortic root up to the left margin of the left subclavian artery and included the aortic arch. The descending aorta corresponded to the segment from the left lateral wall of the subclavian artery up to one image above the origin of the celiac trunk.

Finally, according to the aortic measurements, the studies from both datasets were further labeled as positive or negative for aneurysm(s). Studies were classified as TAA-positive when the ascending diameter was equal to or larger than 4 cm and/or the descending diameter was equal to or larger than 3 cm, regardless of whether the diameters were obtained from a report or were measured on the images.

Automatic System Components

To attain a fully automatic workflow, four components were developed and validated: (a) a series selection filter to automatically select the DICOM series with the greatest spatial resolution for aortic evaluation, (b) a three-dimensional (3D) segmentation model (12) to segment the thoracic aorta from the selected series (with a variable number of images), (c) a postprocessing algorithm to extract the largest diameters from each aortic segment in the segmented volume (13), and (d) a visualization interface of the segmented aortic volume to display the segmented aorta in 3D (14).

DICOM series selection.—A rule-based filter was developed to select the axial series with the greatest spatial resolution from each study to be used as input to the segmentation model. The filter used information from the DICOM header (metadata tags) to perform a selection similar to that performed during the dataset’s annotation and prioritized axial series with the maximum number of image sections and contrast enhancement (when available) in the arterial phase, as determined by the earliest time of acquisition (if multiple contrast-enhanced series were available). The DICOM tags used for filtering included the image reconstruction plane, number of sections, use of contrast material, and the time between contrast material injection and image acquisition.

Thoracic segmentation model.—The segmentation model was based on the 3D U-Net architecture with rectified linear unit activations, 3D batch normalization, 3D spatial dropout, and 3D max pooling layers with size of one in the axial dimension to allow for a variable number of image sections to be input at the inference time. A smoothed Dice similarity coefficient was used as a loss function and was optimized by using the Adam algorithm during training. During inference, the center crop of the image (after isotropic resampling of voxel spacing) was used for segmentation prediction, with overlapping window striding being used to reduce unwanted noise artifacts and edge discontinuities so that overall segmentation quality would be improved. For more details on the data augmentation, model training, and inference, see Appendix E1 (supplement).

Postprocessing diameter extraction.—The postprocessing pipeline consisted of several algorithmic tasks (Fig 2). First, the automatic division of the output segmentation into ascending and descending aortic regions was performed by finding the approximate location of the aortic arch through iteratively scanning down (from the top section) until a section containing two connected components (which should correspond to the ascending and descending aorta) was located; a separating line would then divide these two components (we use the line that connects the midpoint of the two components’ centers of mass to the reciprocal slope). Second, the aortic centerline was determined by using 3D skeletonization, and 100 points along the centerline were sampled after outliers were removed through simple smoothing techniques (median averaging). For each resulting point (a 3D coordinate), the displacement vector from its subsequent neighboring point was calculated such that the segmented region could be rotated in a way that resulted in all displacement vectors being aligned along the same axis (this step is required to orient the plane at each coordinate point so that it is perpendicular to the centerline and the diameter of the aorta can be extracted correctly). Last, an ellipse is fitted around the segmentation mask at each coordinate point to measure the diameter, which is taken as an average of the major and minor axes of the ellipse; for segmented vessels with N diameter measurements, the extracted diameters will be smoothed by applying the following function (f) (R^{N × 2} → R^N) to aggregate the major and minor axes, x₁ and x₂, at each coordinate point:

Figure 2:

Main postprocessing steps performed in a study with an ascending thoracic aneurysm measuring 4.2 cm (true diameter). By using the (A) three-dimensional volume output by the U-Net, the postprocessing algorithm (B) sampled 100 points along the centerline of the predicted volume and (C) extracted the diameters for each point. The x-axis of the plot represents the proximity of each sampled point to the aortic root (in centimeters), and the y-axis represents the predicted diameters for each point (in centimeters).

We considered a prediction to be positive for our application use case of TAA detection if the largest diameter was equal to or larger than 4 cm for the ascending thoracic aorta or if the largest diameter was equal to or larger than 3 cm for the descending thoracic aorta (Figure E1; Movie [supplement]).

Movie:

Download video file ^{(21.3MB, mp4)}

Rotation of a three-dimensional segmentation mask output by the U-Net in different angles.

Visualization interface.—The system also generated a 3D visual output of the segmented aortic volume by using the Visualization Toolkit (https://vtk.org/). The resultant visualization provides the user with reassurance about the model’s prediction quality and allows for greater interpretability of the model’s results by enabling visual inspection of the whole thoracic aorta to confirm the diagnosis (Fig 3). More details can be found in Appendix E1 (supplement).

Figure 3:

Side-by-side comparison between the three-dimensional visualization output by the visualization interface (left) and the volume render (right) in (A) a contrast-enhanced chest CT study without an aneurysm and (B) a contrast-enhanced CT study with an ascending aortic aneurysm. The output is generated through the compilation of a series of Digital Imaging and Communications in Medicine Secondary Capture images containing three-dimensional views of the thoracic aorta so that it can run in a two-dimensional viewer, which is a limitation of some of the commercially available viewers that integrate with picture archiving and communication systems.

Statistical Analysis

For dataset A, the aortic segmentation quality was evaluated by using the Dice similarity coefficient; for dataset B, the quality of 3D output was assessed by visual comparison between the reconstructed thoracic and the corresponding volume render.

The diameter prediction was investigated for both datasets by calculating each aortic segment’s mean absolute error (MAE). Intermethod and interreader agreement (dataset B) were evaluated by using the intraclass correlation coefficient (ICC) for each aortic segment. Bland-Altman plots were created to visualize the intermethod variability of maximal aortic diameter measurements for both datasets and the variability between the two readers for dataset B (interreader bias). Aneurysm prediction was assessed by using the classification metrics of sensitivity, specificity, and accuracy. For each metric, we provided the CIs, which were computed by using bootstrapping. Finally, differences in errors between subsets of dataset B were evaluated by using unpaired independent t tests (series with and without contrast material and series with a section thickness greater than or less than 2 mm) and an analysis of variance test (differences in errors between scanner manufacturers). A P value less than .05 indicated a statistically significant difference. Python version 3.9.6 and RStudio version 1.4.1103 (2009–2021) were used for analysis.

Results

Test Set of Dataset A

The model failed to run properly in 10 of the 155 series from the test set partition of dataset A (failure rate of 6.45%). The Dice similarity coefficient measured between the predicted and the manual reference segmentation in 50 series (48% positive for aneurysm) of the primary test set was 0.92 (Fig 4). The MAE, the ICC, and the difference between the predicted diameters and the diameters from the report are given in Table 3 and in the Bland-Altman plots of Figure 5A. The model performance for identifying aneurysms is reported in Table 4.

Figure 4:

Images of the artificial intelligence model segmentation output (red circles) overlaying the (A) axial CT images of a non–contrast-enhanced CT study with an aneurysm in the ascending aorta, (B) a contrast-enhanced CT study with an aneurysm in the ascending aorta, (C) a contrast-enhanced CT with an aneurysm in the ascending aorta, and (D) a contrast-enhanced CT with dissection in the descending aorta.

Table 3:

Calculated Errors and Interreader Agreement

Figure 5:

Bland-Altman plots of the difference between the predicted and reported diameters for (A) dataset A and (B) dataset B. The x-axis represents the mean of the two measurements, and the y-axis represents the difference between the two measurements. The continuous black line represents the mean value of the differences, and the black dashed lines represent the limits of agreement (±1.96 standard deviation of the mean difference).

Table 4:

Model Performance Results for Aneurysm Classification

Dataset B

Series selection performance.—The system failed to run properly in 11 of the 1400 studies from dataset B (0.79% failure rate). In one of these studies, the automatic series selection failed to run, and in the other 10, the AI model did not run. The automatically selected series matched the series chosen by reader 1 in 1317 studies (94.1%) and matched the series chosen by reader 2 in 1332 studies (95.1%). The series selected by reader 1 matched the ones chosen by reader 2 in 1334 studies (95.3%). More details about the series selected by the system are provided in Table 2.

Measurement performance.—The calculated MAE and ICC referred mainly to TAA-positive studies when the diameters in the report were used as a reference (Table 3 and Fig 5B). Because dataset B comprised routine CT studies, a healthy aorta did not have a diametral measurement in the radiology report most of the time. The MAE and ICC referred to both TAA-positive and TAA-negative studies when the diameters measured by readers 1 and 2 were used as references (Table 3 and Fig 6A, 6B).

Figure 6:

Bland-Altman plots of the difference between the predicted diameters for dataset B and the diameters measured by (A) reader 1 and (B) reader 2 and (C) between the diameters measured by readers 1 and 2. The x-axis shows the mean of the two measurements, and the y-axis represents the difference between the two measurements. The continuous black line represents the mean value of the differences, and the black dashed lines represent the limits of agreement (±1.96 standard deviation of the mean difference).

Estimated errors for the dataset B subsets can be found in Table E1 (supplement). The absolute difference between the predicted diameter and the measured diameter was smaller in the studies with contrast enhancement than in those without contrast enhancement for the ascending aorta when the reference diameter was from reader 1 (P = .004) or reader 2 (P = .02). This difference was also smaller in the studies with contrast enhancement for the descending aorta when the reference was the diameter from reader 2 (P < .001). The absolute difference between the predicted diameter and the measured diameter was also smaller in series with a section thickness smaller than 2 mm than in series with a section thickness equal to or greater than 2 mm for the descending aorta when the reference was the diameter from reader 1 (P = .03) or reader 2 (P = .04). No evidence of a difference was found in the absolute errors between vendors for the ascending thoracic aorta (P = .45 for reader 1 as the reference and P = .16 for reader 2 as the reference) or descending thoracic aorta (P = .82 for reader 1 and P = .58 for reader 2).

Intermethod and interreader agreement.—The intermethod agreement (obtained with manual and automatic methods) and the interreader agreement (between the two readers) are displayed in Table 3. The ICCs between the measured diameter and the predicted diameter for the ascending aorta were 0.69 (95% CI: 0.63, 0.73) when the reference was from the report, 0.81 (95% CI: 0.77, 0.84) when the reference was from reader 1, and 0.82 (95% CI: 0.76, 0.86) when the reference was from reader 2. The ICCs between these diameters for the descending aorta were 0.48 (95% CI: 0.26, 0.64), 0.70 (95% CI: 0.68, 0.72), and 0.75 (95% CI: 0.70, 0.78) when the reference was from the report, reader 1, and reader 2, respectively.

The interreader agreement was strong between readers 1 and 2 for the ascending aorta (ICC, 0.91; 95% CI: 0.90, 0.92) and the descending aorta (ICC, 0.82; 95% CI: 0.80, 0.84). The Bland-Altman plot also supported the results of ICC analysis (Figs 5B, 6). Additionally, the median error between the annotated point on the sample of 100 studies to evaluate the automatic division of the ascending aorta from the descending aorta in the intermediate step of postprocessing was 4 cm.

Classification performance.—Table 4 shows the system’s performance in classifying the presence or absence of aneurysm(s) when using the diameters from the report and from both readers as a reference. The sensitivity and specificity were higher for the ascending aorta, independently of which reference were used. In addition, higher accuracy was observed when the predicted measures were compared with the diameters from reader 2.

Discussion

This work presents a deep learning–based system that uses automatic segmentation of the thoracic aorta to determine aortic diameter measurements on chest CT studies and accurately predict TAAs. In dataset B, 94%–95.1% of the DICOM series automatically selected for TAA detection matched those selected by a human reader (reader 1 or 2), and this performance was comparable with the variability between reader 1 and reader 2 with regard to series selection. For the ascending aorta, the lowest intermethod MAEs were the ones between the automatic diameters and those from the report (MAE, 0.25 cm; SD, 0.38) or from reader 1 (MAE, 0.25 cm; SD, 0.39). For the descending aorta, the lowest intermethod MAE was the one between the predicted diameters and those measured by reader 2 (MAE, 0.26 cm; SD, 0.50). There was higher agreement between the automatic diameters and those measured by the readers than between the automatic diameters and those in the report, and there was also higher agreement between the report’s diameters and the diameters from reader 2 than between the report’s diameters and those from reader 1. Finally, the accuracy of our system for aneurysm classification in the ascending and descending aortic regions were higher when the reference was from either reader 1 (88%; 95% CI: 86, 90 and 81%, 95% CI: 79, 83) or reader 2 (90%; 95% CI: 88, 0.91 and 82%; 95% CI: 0.80, 0.84) than when the reference was from the report (86%; 95% CI: 84, 88 and 66%; 95% CI: 64, 69).

For the ascending aorta, the intermethod and interreader MAEs found in our study were larger than the previously reported MAEs between readers for CT angiograms (0.09–0.13 cm) or between two methods (ie, double oblique views and centerline analysis [0.09–0.20 cm]) (15). However, they were lower than the error described between modalities (0.34 cm) (ie, CT and echocardiography) (16). For the descending aorta, the MAEs found in our study were smaller than those previously reported between two radiologists (0.28 cm) (17). Additionally, as the aortic wall is usually between 0.1 and 0.3 cm thick, the measured aortic dimension can vary by 0.2–0.6 cm, depending on whether both sides of the wall are included between the cursors (18). This is a significant source of variability in measurement and is not considered while assessing the MAE.

Another work that presented a fully automatic (non–AI-based) method to measure the diameters of the ascending and descending aortic regions in multiple locations on non–contrast-enhanced CT studies reported a per-level ICC between manual diameters and automatic diameters ranging from 0.87 to 0.94 (19). However, the diameters produced from their method were only measured at 13 fixed levels relative to the pulmonary arterial bifurcation level (landmark), whereas our method dynamically extracts 100 nonfixed points. Additionally, their method was only capable of segmenting the aorta from 3 cm below the landmark for the ascending aorta to 6 cm underneath the landmark for the descending aorta, whereas our method is capable of intaking the entire length of the thoracic aorta.

Existing semiautomatic and automatic software packages built on traditional tools, including edge detection and other mathematical models, can detect the thoracic aortic centerline and extract its diameters from angiographic CT (20–24) and non–contrast-enhanced CT (25–27). Although non-AI methods (including the commercially available ones) reduce the reporting time and the measurements’ variability, they are often complex and poorly generalizable, require some level of user interaction, and require a substantial level of experience in image postprocessing (28). Unlike previous methods, deep learning methods rely on automatic extraction of hierarchic feature representation directly from image modalities and can use CT images as inputs to output a corresponding segmentation volume (28,29). To the best of our knowledge, there is only one other study that uses an AI-based technique to automatically quantify aortic diameters. In this prior study, the diameters were obtained from nine fixed anatomic landmark positions (regardless of whether a possible aneurysm’s largest diameter corresponded to any of them) from the inner edge to the inner edge of the vessel (without including the aortic wall thickness), and the algorithm was tested in fewer than 20 studies (30).

This study had several limitations. First, the study was retrospective, and the cohort chosen has a greater aneurysm prevalence than the overall population, which may indicate representation bias. If we replicated the reported prevalence of TAAs (estimated between 0.16% and 0.34%) (3), the sample would be heavily skewed toward TAA-negative cases, and as a result, aneurysms involving only the descending aorta would be severely underrepresented. Second, there were limitations related to the annotation tool, such as the lack of a constant distance between the contour points, which could have impacted segmentation accuracy. Finally, we used data from only a single medical center. Although we included different scanners and a large variety of scanned volumes, image-to-noise ratios, and other acquisition parameters, different scanning protocols may require adjustments to the presented system. Analysis of our results provides the basis for additional investigations into potential performance improvements that could be achieved by using semiautomatic tools to speed up aortic segmentation and measurement, incorporating automatic detection of landmarks (such as the right and left subclavian arteries) to delimitate the aortic arch, differentiating the blood lumen from vessel wall diameters, and, finally separating aneurysms from dissection.

In conclusion, the system presented here shows extremely promising results for triaging TAA-positive studies, preventing underdiagnosis in studies requested for other clinical purposes, and providing a more reproducible assessment of thoracic aortic diameters than is currently reported at routine chest CT. The design of this system is also compatible with further integration of a picture archiving and communication system, with autopopulation of results into the radiologist’s report, which is not included in this article.