CT-based True- and False-Lumen Segmentation in Type B Aortic Dissection Using Machine Learning

Abstract

Purpose

To develop a segmentation pipeline for segmentation of aortic dissection CT angiograms into true and false lumina on multiplanar reformations (MPRs) perpendicular to the aortic centerline and derive quantitative morphologic features, specifically aortic diameter and true- or false-lumen cross-sectional area.

Materials and Methods

An automated segmentation pipeline including two convolutional neural network (CNN) segmentation algorithms was developed. The algorithm derives the aortic centerline, generates MPRs orthogonal to the centerline, and segments the true and false lumina. A total of 153 CT angiograms obtained from 45 retrospectively identified patients (mean age, 50 years; range, 22–79 years) were used to train (n = 103), validate (n = 22), and test (n = 28) the CNN pipeline. Accuracy was evaluated by using the Dice similarity coefficient (DSC). Segmentations were then used to derive the maximal diameter of test-set patients and cross-sectional area profiles of the true and false lumina.

Results

The segmentation pipeline yielded a mean DSC of 0.873 ± 0.056 for the true lumina and 0.894 ± 0.040 for the false lumina of test-set cases. Automated maximal diameter measurements correlated well with manual measurements (R² = 0.95). Profiles of cross-sectional diameter, true-lumen area, and false-lumen area over several follow-up examinations were derived.

Conclusion

A segmentation pipeline was used to accurately identify true and false lumina on CT angiograms of aortic dissection. These segmentations can be used to obtain diameter and other morphologic parameters for surveillance and risk stratification.

Supplemental material is available for this article.

© RSNA, 2020

Summary

The authors developed a five-step segmentation pipeline that segments the true and false lumina on CT angiograms in patients with type B aortic dissection and can be used to derive quantitative morphologic features.

Key Points

• ■ A segmentation pipeline was developed that derives multiplanar reformations orthogonal to the aortic centerline from CT angiograms in patients with type B aortic dissection and then segments the true and false lumina; the Dice similarity coefficient was used to assess accuracy and ranged from 0.87 to 0.90.

• ■ The results of the segmentation pipeline were used to accurately and automatically derive maximal diameter in addition to true- or false-lumen cross-sectional area profiles from CT angiograms in patients with type B aortic dissection.

Introduction

Although medical management followed by lifelong clinical and imaging surveillance of aortic diameters is the standard of care in patients with initially uncomplicated type B aortic dissection, it is increasingly recognized that additional morphologic features related to the false and true lumina have important therapeutic and prognostic significance (1–4). Examples include the absolute and relative extent of delamination of the false-lumen outer aortic wall; the number, size, and location of primary intimal tears and exit tears; the orientation of the false lumen and the arrangement of branch vessel origins relative to the false lumen; and the true-lumen shape (5). These features may be helpful for identifying patients at risk for the development of adverse events (eg, aneurysm formation) in the later stages of the disease and those who may thus benefit from early endograft placement.

Assessment and quantification of any of these potentially important features rests on the accurate identification and labeling of the true versus false lumen at imaging. Automated and accurate segmentation of the true and false lumen in patients with uncomplicated aortic dissection would also have an immediate clinical use for measuring aortic diameters, both at baseline and at each follow-up encounter during lifelong surveillance.

Segmentation of aortic dissection from CT images is challenging: Manual segmentation is prohibitively time-consuming for clinical purposes. Traditional computer-vision techniques are poorly suited to deal with the unequal and variable contrast material opacification of the true and false lumina, which may also contain variable amounts of false-lumen thrombus. Segmentation methods that use machine learning methods, specifically convolutional neural network (CNN) architectures, have enabled massive improvements in automated segmentation accuracy (6). CNNs are a specialized type of neural network with numerous layers that can be used for identifying complex visual features relevant for image classification and segmentation tasks (7). Given enough training images encompassing numerous varieties of aortic dissection morphologic features, such algorithms would be expected to become more robust to the variable appearance of aortic dissection than traditional methods.

In this work, we introduce a stepwise CNN-based approach for segmentation that automatically generates multiplanar reformations (MPRs) of the aorta orthogonal to the centerline and then segments the true and false lumina on MPRs. We compare the results of our algorithm to expert-derived manual segmentation. Finally, we demonstrate a clinical use scenario and application of our algorithm by automatically deriving maximum diameters in addition to diameter and cross-sectional area profiles of the aorta in patients with uncomplicated type B aortic dissection undergoing surveillance imaging.

Materials and Methods

Study Population, Imaging Data, and Reference Standard Generation

Our study population includes 45 retrospectively identified patients with initially uncomplicated type B aortic dissection who were followed up for a median of 4 years (range, 10 days to 10.7 years) at a single institution. All patients were selected from an existing cohort of retrospectively identified patients with acute uncomplicated type B aortic dissection from a single institution, as previously reported (8). That study investigated morphologic risk factors for late complications of type B aortic dissection, whereas in this current study, the patients’ CT angiograms were used for the purpose of developing the segmentation pipeline. The study protocol was approved by the institutional review board. The requirement for informed consent was waived because of the retrospective nature of the study.

Patients underwent a total of 153 thoracoabdominal CT angiographic examinations from 2003 to 2017, with a median of three studies and a range of one to nine studies per patient (Table 1). All CT angiographic examinations were performed with state-of-the art multidetector CT equipment, with transverse images reconstructed using a 512 × 512 matrix with pixel spacing of 0.53–1.05 mm and slice spacing of 0.4–3.0 mm. Patients were randomly assigned to a training set consisting of 103 CT angiograms from 30 patients, a validation set consisting of 22 studies from seven patients, and a test set consisting of 28 studies from eight patients (Table 1). There was no overlap in the patients used in the training, validation, and test sets. The same training, validation, and test sets were used for development of the aortic localization (ALO) and true- or false-lumen segmentation (TFL) algorithms described below. The same validation and test sets were used for testing of the full segmentation pipeline.

Table 1:

Training, Validation, and Test-Set Characteristics

All 153 CT angiographic data sets were manually segmented with software (Intuition; TeraRecon, Foster City, Calif) by four experienced radiologists (M.J.W., with 8 years of experience in radiology; K.H., with 8 years of experience in radiology; M.K., with 5 years of experience in radiology; and L.D.H., with 2 years of experience in cardiovascular imaging) into “true lumen,” “false lumen,” and “background” labels on axial slices to generate a reference standard for segmentation. If a thrombus was present in the false lumen, it was included and labeled as a false lumen. Manual segmentation took approximately 1–2 hours per data set, depending on the quality of the scan and the complexity of findings.

Segmentation pipeline development.—Our segmentation pipeline consists of five steps (Fig 1). First, a deep learning algorithm segments the entire aorta on axial slices. Second, the resulting aortic volume is used to determine the path of the aortic centerline. Third, MPRs orthogonal to the centerline are generated. Fourth, a second deep learning segmentation algorithm is applied to the MPRs for segmentation of the aorta into true and false lumina. Finally, the MPR segmentation is optionally converted back to the axial plane; for applications such as derivation of aortic diameter, this step is unnecessary. The steps are explained in detail below.

Figure 1:

Segmentation pipeline. 0, source image. 1, An initial segmentation algorithm that accepts axial images is used to segment and localize the aortic lumen (aortic localization [ALO] algorithm) as a whole. 2, The resulting segmentation is used to derive the aortic centerline (CEN algorithm). 3, The centerline is used to generate multiplanar reformations (MPRs) orthogonal to the centerline. 4, The true or false lumen segmentation (TFL) algorithm is used to segment the true lumen, false lumen, and background. 5, Segmentation in the MPR plane is optionally converted (CON) back to the axial plane. Green = aortic lumen, blue = true lumen, red = false lumen.

In step 1, the ALO segmentation algorithm takes axial CT angiogram slices and segments voxels into “aorta” or “background” labels using a deep learning–based algorithm. The underlying algorithm architecture was TernausNet (9). TernausNet is similar to other encoder-decoder network architectures that are now commonly used for segmentation, but it uses VGG encoder blocks (10) that were pretrained on ImageNet. For training and testing purposes, the “aorta” labels were derived from the manually segmented data by taking the union of “true lumen” and “false lumen.” In total, there were 76 638 images in the training set, 16 709 in the validation set, and 19 672 in the test set (Table 1). Augmentation methods included contrast, brightness, and rotation alterations. The network was implemented in PyTorch (https://pytorch.org/) and trained over 15 epochs on a workstation with two NVIDIA GeForce GTX 1080 Ti graphics cards (NVIDIA, Santa Clara, Calif) with 12 GB of RAM, an Intel Core i7–8700 K central processing unit at 3.7 GHz (Intel, Santa Clara, Calif), and 32 GB of RAM. The loss function was binary cross-entropy. The initial learning rate was set at 0.0001, and the Adam optimizer function was used. The batch size was eight.

The ALO algorithm was applied to axial images from the validation set. The Dice similarity coefficient (DSC) was used to assess performance and is defined as two times the number of elements in the intersection of two sets divided by the sum of the number of elements in each set. The DSC was calculated for the entire aortic volume (in which the sets were the voxels of the predicted vs ground truth segmentation). Following tuning of hyperparameters, the ALO algorithm was tested on the independent test set.

In addition, the average Euclidean distance between edges of the predicted versus ground truth segmentation was assessed. Specifically, the distance between the closest voxel along the ground truth segmentation boundary to the predicted segmentation boundary within each slice was computed and the average distance for each slice determined. Then, the average Euclidean distance across all slices containing the aortic segmentation was computed for each study. Finally, the means and standard deviations of the average Euclidian distances across all studies were computed.

In step 2, the centerline of the aorta was extracted from the three-dimensional segmentation mask derived from the ALO algorithm by using a sequential thinning skeletonization technique (11). As each of these skeletons contains many spurious side branches, the centerline was identified by determining the longest path through the aorta. To ensure a smooth curve, the centerline was then approximated with three-dimensional cubic B-splines.

For assessment, aortic centerlines were also extracted from reference expert–segmented data sets, and the centerline positions and positions relative to the aorta were visually compared with the automated centerlines. In addition, two-dimensional MPRs were generated relative to the aortic centerline (described below) and cross-sections along the aorta were compared to confirm that there was no substantial obliquity to the MPRs.

In step 3, the smoothed centerline was then used to resample each CT angiographic data set and generate MPRs orthogonal to the centerline. To minimize rotation of the frame on consecutive slices, an algorithm employing rotation-minimizing frames was applied (12). For each CT angiogram, 500 MPRs orthogonal to the aortic centerline were generated automatically from the aortic root to the aortic bifurcation.

In step 4, the TFL segmentation algorithm takes MPRs of the aorta as input and segments voxels into “true lumen,” “false lumen,” and “background” labels using a deep learning algorithm.

The TernausNet architecture for segmentation was again used. Because of the multiclass nature of this segmentation step, the loss function was selected to be binary cross-entropy minus log Jaccard.

The ground truth segmentations were used in combination with the centerline extraction and MPR generation methods (steps 2 and 3) to generate 500 evenly spaced MPR images orthogonal to the centerline for each CT angiogram. Hence, in total, there were 51 500 MPR images in the training set, 11 000 in the validation set, and 14 000 in the test set (Table 1). All images were 256 × 256 in size. Relative to the original data set, a 2× upsampling was performed such that pixel spacing was half that of the original axial slices. Most training parameters were the same as those used in the ALO algorithm, including learning rate, batch size, and number of training epochs. There were small differences in the frequency with which augmentation options were used.

The TFL algorithm was tested by applying the algorithm to MPR images generated from the ground truth centerline for CT angiograms in the validation set. The DSC was calculated on a per-case basis for the true lumen, false lumen, and entire aorta across the entire stack of resampled slices by comparing the predicted and ground truth labels of voxels. The “entire aorta” label was determined by taking the union of all voxels classified as “true lumen” or “false lumen.” Following optimization, the TFL algorithm was tested on the independent test set. The Euclidean distance between lumen boundaries was also computed as described earlier for the ALO algorithm.

For step 5, results of the segmentation in the MPR slices can be used to relabel voxels of the original axial data set by transforming MPR image labels back to the original axial slices. Although this step is not necessary for aortic diameter, circumference, or cross-sectional area measurements, it would be important for future applications, such as volumetric calculations.

Full segmentation pipeline testing.—After the individual components of the segmentation pipeline were developed, the fully automated segmentation pipeline was tested on the validation and test set CT angiograms without additional training. The input to the full pipeline was axial CT angiogram slices (step 1). The ALO algorithm was first applied to generate an initial segmentation of the aorta on axial slices. The resulting volume was then used as the input to the centerline-generation algorithm and MPR-generation algorithm steps to automatically generate a centerline and 500 MPRs orthogonal to the centerline (steps 2 and 3). Finally, the TFL algorithm was applied to the orthogonal MPRs (step 4), the output of which was segmentations of MPRs into the true lumen, false lumen, and background.

The DSC was calculated for the true lumen, false lumen, and the entire aorta across the entire stack of resampled slices for each validation and test set case. Ground truth segmentation for the MPR slices was determined by transforming the axial ground truth segmentations into the automatically generated MPR plane based on the automated centerline; hence, the exact slices slightly differed from those used to develop the TFL algorithm because of minimal differences between the centerlines based on manual versus automated segmentation of the aorta on axial slices. The Euclidean distance between lumen boundaries was also computed as described earlier for the ALO algorithm. To measure the frequency of true- and false-lumen mislabeling, we determined the percentage of slices in which the reversal of true- and false-lumen labels led to improved true- and false-lumen DSCs.

Derivation of quantitative morphologic parameters for surveillance.—Clinical application of true- and false-lumen segmentation is not based on voxel-per-voxel accuracy but is rather based on detection and measurements of features such as diameters for true or false lumina and, most important, the entire aortic diameter. Therefore, we derived several parameters that are currently used in our clinical practice (maximum diameter, diameter profiles) in the setting of aortic dissection surveillance. We also show that cross-sectional area parameters can also be extracted from the segmented data sets.

Automatic derivation of maximum diameter, true-lumen area, and false-lumen area was implemented in Python by using OpenCV (https://opencv.org/). Diameter was calculated for all 500 automatically generated MPRs for each test set CT angiogram. A median filter was applied to the diameter profile to smooth the distribution. For diameter and area plots, manual inspection of consecutive MPR images was used to determine the slices corresponding to the right coronary artery takeoff and the aortic bifurcation; this enabled direct comparison of diameter profiles across the same positions between follow-up studies. Manual maximal size measurements, similar to those used in our clinical aortic surveillance program, were performed by using Intuition software to serve as a reference standard.

Manual and automated diameters were compared graphically, and the coefficients of determination (R²) were calculated.

Statistical Analysis

For the ALO algorithm, TFL algorithm, and full pipeline segmentations, determination of statistical differences between validation and test sets was performed with the Mann-Whitney U test. For the TFL algorithm and full pipeline testing, comparison of the DSCs and Euclidean distances among the true lumen, false lumen, and total aorta was performed by using the Wilcoxon signed rank test. In addition, a Bland-Altman plot was created.

Results

Segmentation Pipeline Development

Aorta localization (ALO algorithm).—Highly accurate segmentation of the aorta was obtained on the basis of axial CT angiogram slices by using the ALO algorithm, with an average DSC (±standard deviation) of 0.952 ± 0.010 for the validation set and 0.946 ± 0.011 for the test set (Table 2). The mean Euclidean distance between the ground truth and predicted segmentation boundary was 1.59 mm ± 0.62 for the validation set and 1.36 mm ± 0.96 for the test set..

Table 2:

Accuracy of Segmentation Algorithms for Aortic Dissection

Centerline and MPR generation (centerline and MPR algorithms).—Visual inspection of centerlines demonstrated that centerlines automatically derived from the ground truth segmentation were similar to those derived automatically. Similarly, MPRs derived by using the ground truth segmentation appeared visually similar to those derived from automatic segmentation.

True- and false-lumen segmentation (TFL algorithm).—Using centerlines derived from the ground truth segmentation, MPRs were generated for validation and test-set data. The average DSCs for true and false lumina were high in both the validation and test sets, ranging from 0.884 to 0.906 (Table 2), but were less than those obtained for the entire aorta. The mean Euclidean distance between the ground truth and predicted segmentation boundary was less than 2 mm for all categories (range, 0.77–1.32 mm).

Full Segmentation Pipeline Testing

The full segmentation pipeline was applied to the axial CT angiograms of the validation and test sets to obtain automated segmentation into the true lumen, false lumen, and background on MPRs orthogonal to automatically derived centerlines. The average DSC was high, ranging from 0.873 to 0.900 for the validation and test sets, and accuracy was similar to that obtained by applying the TFL algorithm to MPRs on the basis of ground truth centerlines. The average Euclidean distance between the ground truth and predicted segmentation was less than 3 mm for all categories (mean, 0.83–2.69 mm).

Visual inspection showed that segmentation of most MPRs was highly accurate (Fig 2, Movie [supplement]) and robust to scenarios including false-lumen thrombosis, poorly opacified false lumina, and the presence of branch vessels. In relatively rare cases, there were errors of segmentation in which true and false lumina were reversed or the border of the lumen was incorrectly identified (Fig 3). The number of slices in which there was mislabeling of true and false lumina was calculated to occur in 3.8% of the MPR slices of the validation set and 3.3% of the slices of the test set. Although reasons for incorrect classification were not apparent in all cases, errors tended to occur in areas of increased image noise with decreased visualization of the dissection flap, which in turn tended to occur more often in the abdominal aorta. Invariably, the algorithm-generated segmentation resulted in smoother contours of the true and false lumina when compared with the manual segmentation because of the increased resolution of the MPRs compared with the original images.

Figure 2:

Segmentation of multiplanar reformations obtained by using the fully automated pipeline. Segmentation was robust to the presence of false-lumen thrombosis (top), partial opacification of the false lumen (middle), and the presence of branch vessels (bottom).

Figure 3:

Examples of segmentation misclassification. In some multiplanar reformation images, there was gross misclassification of true and false lumina (top). In other images, the borders of the true and false lumina were inaccurately detected (bottom).

Movie 1:

Download video file ^{(16.5MB, mp4)}

Segmentation results on multiplanar reformations (MPRs) along the aortic centerline using the fully automated pipeline for a test set patient. Source MPR image is shown on the left. Ground truth segmentation is shown in the center. Predicted segmentation is shown on the right. Gray: true lumen. White: false lumen.

Total processing time varied from 1.5 to 4 minutes per study, with approximately 75%–80% of this time corresponding to centerline extraction and conversion of axial images to MPR images. By comparison, most manual segmentations took approximately 1–2 hours.

Derivation of Quantitative Morphologic Parameters for Surveillance

Automatically derived maximal diameters for eight baseline and 20 follow-up CT angiograms from all 28 patients in the test set were correlated with manual maximal diameters, resulting in an R² of 0.95 (Fig 4, A) and an average error of 0.1 mm ± 1.6. Automated measurements never differed from manual measurements by more than 3 mm (Fig 4, B). There was no correlation between the maximal diameter and margin of error.

Figure 4:

Maximal diameter measurements for test set CT angiograms. A, Scatterplot of manual maximal diameter measurement shows good correlation with automated maximal diameter measurement. The line of best fit with 95% confidence interval is displayed. Data in parentheses are 95% confidence intervals. B, Bland-Altman plot. The difference between manual measurement and automated measurement was less than 3 mm for all examples; the magnitude of difference did not correlate with maximal diameter. SD = standard deviation.

Automated segmentations were used to generate diameter plots along the length of the aorta for patients in the test set as an illustration of how this information might be used in a clinical setting; a representative example including a patient’s baseline CT angiogram and three follow-up studies is shown in Figure 5. Automatic diameters correlated well with data derived from ground truth segmentation. Areas of growth, such as the proximal descending aorta, are apparent in these plots. Although agreement between predicted and ground truth diameter was generally strong, there was increased disagreement between the predicted and ground truth diameter in some specific regions (eg, the suprarenal abdominal aorta in the baseline examination and the infrarenal abdominal aorta on the third follow-up of this patient). Inspection of these studies demonstrated higher image noise in these locations, which likely affected the quality of segmentation.

Figure 5:

Aortic diameter plots generated from automated segmentations versus ground truth segmentation for a patient with three follow-up CT angiographic studies. Good correlation is seen between the diameter profile based on the ground truth segmentation versus automated segmentation. Areas of growth and areas that are unchanged over time are apparent.

Similar profiles were generated for the cross-sectional area of the true lumen, false lumen, and entire aorta, which also showed good correlation between automated and ground truth data (Fig 6).

Figure 6:

A, True-lumen, B, false-lumen, and, C, total cross-sectional area plots generated from automated segmentations versus ground truth segmentation for a patient with three follow-up CT angiographic studies. Good correlation is seen between the profiles based on the ground truth segmentation versus automated segmentation.

Discussion

In this work, we developed a stepwise segmentation pipeline that includes two deep learning algorithms to automatically identify the true and false lumina in patients with uncomplicated type B aortic dissection. We have shown that our method is reasonably accurate compared with expert manual segmentation, and we have demonstrated its ability to derive automated measurements in a clinical use scenario of aortic surveillance imaging.

Identification and labeling of the true and false lumina is a prerequisite for the extraction of a wide range of morphologic features that have recently been shown to be predictive of adverse events (2,3,5). Understanding such risk factors may ultimately determine which patients should undergo endovascular repair, a crucial determination given the potential mortality benefits weighed against significant procedural risks (4,13,14). Accurate segmentation of the true and false lumina is also needed for fluid dynamic simulations (15) and may also improve our understanding of the complex hemodynamics over different phases of the disease (16).

Although we have primarily performed this work as a first step toward deriving morphologic risk factors for late complications in type B aortic dissection, a natural use of this technique was to generate diameter profiles for potential integration into the workflow of our current aortic surveillance program. We were able to obtain maximal diameters within several millimeters of manual measurements, close to the limits of manual measurement error (17). Existing clinical workflows typically obtain aortic diameter measurements at discrete anatomic landmarks that may fail to capture subtle growth trends; in our study, superimposing diameter measurements obtained at different time points into a single plot generated an intuitive, comprehensive visualization of aortic growth over time that would likely be valuable to physicians caring for these patients (18,19).

A deliberate decision in the design of our segmentation pipeline was to divide the final segmentation into true and false lumina on the plane orthogonal to the aortic centerline (the MPR plane) rather than on the axial plane. Our intention was to mirror the clinical workflow in which the borders of the aorta are identified on the MPR plane before measurement. We anticipated that using this technique in the clinical workflow would always require visualization of the segmentation on the MPR plane; therefore, rather than performing segmentation in the axial plane first and then transforming segmentations into the MPR plane, which could potentially introduce artifacts related to this transformation, we opted to first generate MPRs and then perform segmentation. We did attempt the former approach as well and found that first converting to the MPR plane led to small improvements in segmentation accuracy (Table E1 [supplement]). We suspect that this at least in part relates to MPR slices providing a more consistent appearance of the aorta, with clearer delineation of the dissection flap when compared with axial slices, particularly along the aortic arch.

We were interested to find that the segmentation algorithm worked well using two-dimensional slices without the aid of information from adjacent proximal or distal slices. This implies that, in most cases, the true lumen can be visually distinguished from the false lumen on the basis of individual two-dimensional slices by using visual cues, a fact which is already well known to radiologists (20).

Previous efforts in aortic CT segmentation have generally used a staged approach with traditional methods in computer vision. Common techniques include thresholding or anatomic segmentation (21–23) to remove extraneous structures; localization of anatomic structures (22–24) for aortic seed-point identification followed by region-growing methods (22,25); aorta localization through the Hough transform (22,26,27); and atlas-based segmentation (21,23,28). These approaches have been effective for segmentation of the aorta as a whole in both unenhanced and contrast material–enhanced CT, resulting in DSCs in the range of 0.85–0.93 (23,27,29).

Comparatively little work has been performed on segmentation of aortic dissection. A variety of methods have been used to detect the dissection flap, including multiscale template filters (30) and analysis of Hessian matrix eigenvectors (28,31). These efforts have generally been validated on few if any patients (25,30,32,33) and often explicitly assume that the dissection flap is a low-attenuation edge surrounded by a well-opacified lumen on either side. Therefore, it is questionable whether they would be robust to the wide variations of aortic structure in the setting of dissection related to poor lumen opacification or the presence of thrombi.

By contrast, machine learning algorithms based on CNNs should be able to recognize dissection membranes with varying appearances, so long as there is sufficient representation of similar structures in the training set. We were able to confirm this hypothesis by demonstrating successful false-lumen segmentation with varying degrees of contrast opacification and false-lumen thrombosis.

Two other studies have recently used CNNs to perform segmentation of aortic dissection. In the first study, Li et al (34) used CNNs to identify points along the boundary of the aorta and true lumen rather than performing voxel-wise class identification. Fivefold cross-validation on a total of 45 CT scans was performed, and excellent results were reported, with DSCs of 0.989 for the aorta and 0.925 for the true lumen. However, the inputs to the algorithm were MPRs extracted from a centerline of the aorta, which in turn was generated from manual segmentation of the aorta; MPRs were automatically extracted in our study.

The second study by Li and colleagues (35) is more similar to our work, given its use of a large training set of 254 CT angiograms, an encoder-decoder–style algorithm architecture, and a fully automated segmentation pipeline; however, their study importantly used a three-dimensional rather than a two-dimensional CNN architecture. The use of a three-dimensional architecture would be expected to be useful for gathering global spatial information, potentially allowing improved identification of the true lumen.

On the other hand, we believe our pretrained two-dimensional architecture offers several advantages. First, our pretrained architecture allowed us to rapidly optimize hyperparameters; hyperparameter tuning would be expected to take substantially longer if network weights were trained from scratch. Second, our use of slices oriented along the aortic centerline may have resulted in improved segmentation because they result in a more consistent, circular appearance of the aorta centered in each MPR image when compared with transverse images. In addition, this may lend itself to greater improvements from rotational augmentation techniques. Finally, our network architecture requires no downsampling of images (and in fact upsamples the original data before segmentation) because of the smaller number of parameters and corresponding graphics processing unit demands of our two-dimensional architecture.

Although direct comparison with previous studies is impossible given differences in patient populations, input formats, and imaging conditions, the ranges of DSCs obtained in our study and in the two studies employing CNNs were similar.

Our study had some limitations. A potential limitation was the use of multiple studies from individual patients. Although we deliberately did this in anticipation of comparing parameters for individuals over successive examinations, our design could have introduced bias into the validation and test sets by decreasing the variety of aortic dissection structures. Future work will need to be validated in a larger set of distinct patients. In addition, we believe that practical adoption of this pipeline will inevitably require a degree of manual quality control, given the range of potential aortic dissection appearances. Thus, this method will greatly expedite but not replace manual segmentation.

In conclusion, we developed a stepwise, reasonably accurate CNN-based approach for segmentation of CT angiograms of type B aortic dissection into true and false lumina. In the current clinical workflow, this method can be used for surveillance of aortic size. In the future, we envision this method enabling large-scale extraction of morphologic parameters in cohorts of patients with type B aortic dissection, which will expedite the study and discovery of prognostic factors and ultimately guide treatment decisions.