Deep learning approach for the segmentation of aneurysmal ascending aorta

Abstract

Diagnosis of ascending thoracic aortic aneurysm (ATAA) is based on the measurement of the maximum aortic diameter, but size is not a good predictor of the risk of adverse events. There is growing interest in the development of novel image-derived risk strategies to improve patient risk management towards a highly individualized level. In this study, the feasibility and efficacy of deep learning for the automatic segmentation of ATAAs was investigated using UNet, ENet, and ERFNet techniques. Specifically, CT angiography done on 72 patients with ATAAs and different valve morphology (i.e., tricuspid aortic valve, TAV, and bicuspid aortic valve, BAV) were semi-automatically segmented with Mimics software (Materialize NV, Leuven, Belgium), and then used for training of the tested deep learning models. The segmentation performance in terms of accuracy and time inference were compared using several parameters. All deep learning models reported a dice score higher than 88%, suggesting a good agreement between predicted and manual ATAA segmentation. We found that the ENet and UNet are more accurate than ERFNet, with the ENet much faster than UNet. This study demonstrated that deep learning models can rapidly segment and quantify the 3D geometry of ATAAs with high accuracy, thereby facilitating the expansion into clinical workflow of personalized approach to the management of patients with ATAAs.

Introduction

An ascending thoracic aortic aneurysm (ATAA) represents a permanent vessel dilatation of the aorta leading to adverse events and death. Nearly 10 out of 100,000 persons per year are affected by ATAA [1], with the congenital bicuspid aortic valve (BAV) having a reported prevalence of the aortopathy in the range of 20–84% [2]. Indeed, patients with BAV have an 80-fold higher risk of developing an ATAA than the general population with the morphologically-normal tricuspid aortic valve (TAV) [3]. If left untreated, an ATAA can lead to fatal complications such as an aortic rupture or dissection. Elective surgery to avoid aortic complications is indicated when the aortic diameter exceeds 5.5 cm since the yearly risk of dissection or rupture rises from 3 to 7% with aneurysms > 6 cm [4]. Although diagnostic imaging is an essential step to measure the critical aortic diameter of an ATAA, rupture and dissection may occur at aortic size not falling within surgical guidelines so additional metrics not based on size are needed to improve the clinical decision-making process [5]. Manual or semi-automatic segmentation from computerized tomography (CT) or magnetic resonance imaging (MRI) data is usually performed for aortic size evaluation. It is also important to remember that size is not the only important imaging marker; aortic shape matters as well as the loss of the normal “waist” of the aorta at the sinotubular junction.

Deep learning methods are emerging for vascular segmentation and remains a challenging area of research [6–10]. These techniques have shown tremendous success in the last 5 years for image classification and segmentation tasks in various fields, especially for neuroimaging for small vessel segmentation [11]. In general, the deep learning approach requires (1) any medical images (e.g. MRI, or CT images), (2) a volume of interest according to the desired classification output, (3) the training and testing of the deep learning algorithm and (4) the validation of semantic segmentation. Deep learning techniques may appear simpler and more flexible than machine learning but require more quantities of labelled data for the training process and are usually more complex and less transparent (the so-called ‘black box’). This has limited the widespread adoption of deep learning in clinical practice. However, several different categories of deep learning models have been proposed for image segmentation with the UNet being the most adopted technique for biomedical imaging analysis [12]. In a different way, efficient neural network (ENet) [13] and the efficient residual factorized convnet (ERFNet) [14] are commonly used in mobile applications where hardware availability is limited and accurate segmentation is very critical.

This study aims to develop a deep learning framework for the segmentation of the aneurysmal aorta and its valve. Specifically, deep learning vessel segmentation was developed using 72 CT scans of patients with ATAAs and different aortic valve morphology (i.e., BAV and TAV). Three different deep learning models including the UNet, ENet, and ERFNet were investigated to account for accurate vessel segmentation, fast training time, low hardware requirements for inference, and low training data requirements. Cross-validation strategy was applied for training. The three deep learning models were compared to reveal accurate segmentation of ATAAs with a small sample size.

Materials and method

Study population and CT imaging

After internal review board approval (IRRB_04_04 released by Comitato Etico sezionale IRCCS ISMETT) and informed consent, a total of 72 ATAAs collected from patients underwent electrocardiographic-gated computed tomography (ECG-gated CT) angiography for aortic size evaluation were enrolled. For all patients, ECG-gated CT scans were done after intravenous injection of contrast agent to improve image quality. The CT examination was carried out on a GE VCT 64-channel scanner (GE Medical Systems, Milwaukee, Wisconsin) with gantry rotation velocity of 0.5 m/s and spiral pitch of 0.984. This allowed us to obtain 10-phase ECG-gated thoracic data sets of the entire cardiac cycle with a resolution of 512 × 512, and slice thickness of 0.625 mm. In this study, we selected the cardiac phase showing the aortic valve at fully opened shape, which frequently occurs at 50–100 ms after the R peak. Aortic valve morphology (i.e., BAV vs. TAV) was assessed by an experienced radiologist using images reconstructed parallel to the aortic valve plane.

For training of deep learning networks, semi-automatic thresholding of the contrast-enhanced images followed by manual cropping and morphologic operations was performed to generate 3D virtual masks of whole aorta using Mimics software (Materialize NV, Leuven, Belgium). This was performed by a 10-year experienced user as previously done by our group [15–17]. All 3D segmented masks had size characterized by isotropic voxel size of 1 × 1 × 1 mm³ and matrix resolution of 512 × 512. Segmentation masks were resampled using nearest neighbour interpolation and converted to binary values with 0 for background and 1. We then implemented UNet (Sect. 2.2), ENet (Sect. 2.3) and ERFNet (Sect. 2.4) networks and Tversky loss using Keras with Tensorflow (Sect. 2.5) in the open-source mathematical programming language Python (www.python.org).

UNet model

Several changes were made to the original UNet architecture to improve segmentation results [12], as showed in Fig. 1. All 3 × 3 convolutions were replaced by larger 5 × 5 convolution operators. Each convolution was followed by a drop out [18] layer with rate of 10%. Dropout layers help to regularize the network and avoid overfitting. While the original UNet architecture does not use padding when applying convolution operators, we adopted zero padding to ensure that the size of the output feature map is the same as the input size. The original UNet has a 2D size of 32 × 32 along with 1024 feature maps at the final layer of the contraction path. In a different way, we used an input size of 512 × 512 with 32 filters on the first contraction path layer, with doubling of feature maps after each max pool and stopping at 256 feature maps and 2D size of 64 × 64.

Fig. 1.

Comparison between the original UNet architecture (a) [12] and our UNET implementation (b). Each blue box is a multi-channel feature map with the number of channels denoted at the top of the box. The x–y size is denoted at the bottom lower left edge of the box. White boxes represent copied low-resolution features

ENet model

The ENet represents optimized neural network developed for fast inference and high accuracy, which typically occur in augmented reality and automotive [13]. The ENet architecture was based on building blocks of residual networks, with each block consisting of three convolutional layers. These were a 1 × 1 projection that reduces dimensionality, with a regular convolutional layer and a 1 × 1 expansion along with batch normalization and surpassing human-level performance. ENet adopted several types of convolutions to build an encoder/decoder style image segmentation network. In some layers, ENet had asymmetric convolutions characterized by separable convolutions with sequence of 5 × 1 and 1 × 5 convolutions. The 5 × 5 convolution had 25 parameters while the corresponding asymmetric convolution had only 10 parameters to reduce the network size. Finally, the ENet used a single initial block in addition to different variations of the bottleneck layer. Figure 2 shows the ENet architecture [13].

Fig. 2.

ENet architecture [13]. a ENet initial block with 2 × 2 max pooling with a stride of 2 and convolution has 15 filters, summing to 16 feature maps after concatenation. b ENet bottleneck module. ‘conv’ is either a regular, dilated, or full convolution (deconvolution) with 3 × 3 filters, or a 5 × 5 convolution decomposed into two asymmetric (separable) ones

ERFNet-model

Inspired by residual networks and ENet, ERFNet was optimized to improve accuracy and efficiency in image segmentation with respect to ENet [14]. This leads to more accurate segmentations for urban scenario. The basic building block module of ERFNet segmentation network was referred to as a non-bottleneck-1D layer (see Fig. 3) and comprised of two sets of factorized (separable or asymmetric) convolutions of size 3 × 1 followed by the 1 × 3 with rectified linear unit non-linearity. The input feature map of the main convolution path was added element-wise to the output of the convolution path, which represented the input of the next layer after applying the rectified linear unit non-linearity. Size of ERFNet input was 512 × 512 while the down-sampler block was similar to that of ENet architecture. This architecture was based on dilated convolutions with different sizes in the Non-bt-1D layers as well as spatial dropout as regularizer.

Fig. 3.

Basic building block layer of ERFNet network called Non-bottleneck-1D (Non-bt-1D) [14]

Training methodology

Loss function

Deep learning methods generally suffer from imbalanced data problems [19]. This problem is common in biomedical image segmentation where the anatomy of interest may be very small compared to the background consisting of connective tissue with a wide range of intensity grey values. In such imbalanced data problems, the network tends to simply predict most voxels as belonging to the background class. Loss functions can be adopted to solve the class imbalance problem and provide a large weight to foreground voxels.

To overcome imbalanced data problem, we adopted Tversky loss function [20] assuming the dice similarity coefficient (DSC) as:

$$DSC=\frac{2\left|P\cap G\right|}{\left|P\right|+\left|G\right|}$$ 1

where P and G are the set of predicted and ground truth labels, respectively. To make better adjustment of the weights of false positive (FP) and false negative (FN), we adapted a penalty approach as a follow:

$$S\left(P,G;\alpha \beta \right)=\frac{\left|P\cap G\right|}{\left|P\cap G\right|+\alpha \left|P\backslash G\right|+\beta \left|G\backslash P\right|}$$ 2

where $\alpha$ and $\beta$ control the magnitude of penalties of FPs, and FNs and P\G is the relative complement of G on P. Therefore, the Tversky loss function can be defined as:

$$T\left(\alpha \beta \right)=\frac{{\sum }_{i=1}^N{p}_{0i}{g}_{0i}}{{\sum }_{i=1}^N{p}_{0i}{g}_{0i}+\alpha {\sum }_{i=1}^N{p}_{0i}{g}_{1i}+\beta {\sum }_{i=1}^N{p}_{1i}{g}_{0i}}$$ 3

where the output of the final layer of the network (soft-max layer), p_0i is the probability of voxel i to be part of ATAA wall and p_1i is the probability of it belonging to the background. Also, the ground truth training label g_0i is 1 for ATAA wall and 0 for everything else (background) and vice versa for the g_1i. By adjusting the parameters, $\alpha$ and $\beta$, the trade-off can be controlled between FPs and FNs. Setting $\alpha$ = $\beta$ = 0.5 leads to the familiar DSC while setting $\alpha$ + $\beta$ = 1 leads to a set of $F_{\beta }$ scores, ${\beta }^{\prime }$ s larger than 0.5 weight recall higher than precision by placing more emphasis on FNs leading to better segmentation in slices with small foreground area.

Training

In stratified five-fold cross-validation, the data set (72 patients) were divided into 5 equal patient subsets, and the holdout method repeated 5 times. Consequently, for each of the three network models, we trained each single model 5 times. Each time, one of the 5 subsets was used as the testing set, and the other 4 subsets as training set. Slices from the same patient were never used for both training and testing purpose. So, there was no cross-contamination between training and test sets.

Data augmentation was used to train neural network models to reduce overfitting. This was applied by randomly rotating and translating in both x and y directions, and then applying shearing, horizontal flip and zooming to the input training image slices. Additionally, data standardization or normalization was performed as a pre-processing step to prevent the weights from becoming too large and thus avoid numerical instability. For each fold, 2D pixel-wise mean and standard deviation were computed using all training data. Specifically, each patient data was standardized subtracting the mean and dividing by the standard deviation. An initial set of 16 patients for determining the best learning rates for each of three models was used. A learning rate of 0.0001 for ENet model and 0.00001 for ERFNet and UNet models with Adam optimizer were adopted [21]. We also adopted a batch size of 8 slices for all experiments and adopted a Tversky loss function with $\alpha$ = 0.3 and $\beta$ = 0.7 as reported by Salehi et al. [20]. These values for the loss function demonstrated to be effective for training deep neural networks for sclerosis lesion segmentation. All models were trained with a maximum of 100 epochs; particularly, an automatic stopping criterion ending the training step when loss decreased upon 10 epochs was implemented.

A high-end HPC system equipped with a GPU (NVIDIA QUADRO P4000 with 8 GB of RAM) was used to train all networks and run inference.

Data analysis

For each clinical case, sensitivity, positive predictive value (PPV), Dice score (DSC), volume overlap error (VOE), relative volume difference (VD), and average symmetric surface distance (ASSD) were computed to compare the performance of each deep learning network [22, 23].

$$Sensitivity=\frac{\mathit{TP}}{TN+FN}$$ 4

$$PPV=\frac{\mathit{TP}}{TP+FN}$$ 5

$$DSC=\frac{2TP}{2TP+FP+FN}$$ 6

$$VOE=1-\frac{\mathit{TP}}{TP+FP+FN}$$ 7

$$VD=\frac{\left|FN-FP\right|}{2TP+FP+FN}$$ 8

$$ASSD\left(X,Y\right)=\frac{\left\{ASD\left(X,Y\right)+\left(ASD\left(Y,X\right)\right\}\right)}{2}$$ 9

where

$$ASD\left(X,Y\right)=\sum _{xϵX}mi{n}_{yϵY}\frac{d\left(x,y\right)}{\left|X\right|}$$ 10

Analysis of variance (ANOVA) on the DSC was used to assess statistical differences among network. Statistical significance was considered for α ≤ 0.05.

Results

Table 1 shows the performance of ATAA segmentation computed using the ENet, UNet and ERFNet methods. It can be observed that the ENet had the highest DSC module (91.2 ± 8.9%) as compared to those of both UNet (91.1 ± 10.2%) and ERFNet (88.4 ± 9.9%).

Table 1.

Performance segmentation using the ENet, UNet and ERFNet methods

	Sensitivity (%)	PPV (%)	DSC (%)	VOE (%)	VD (%)	ASSD
ENet
Mean	92.69	90.67	91.22	15.22	2.83	4.46
± std	11.13	9.49	8.97	11.62	14.47	4.54
± CI (95%)	2.57	2.19	2.07	2.68	3.34	1.05
UNet
Mean	91.63	91.79	91.09	15.30	0.12	5.48
± std	12.14	7.21	10.18	11.76	16.09	4.67
± CI (95%)	2.80	1.66	2.35	2.72	3.72	1.08
ERFNet
Mean	89.01	88.94	88.41	19.56	0.92	5.48
± std	12.46	10.33	9.94	13.88	16.87	4.67
± CI (95%)	2.88	2.39	2.30	3.21	3.90	1.08

At analysis of variance, the P value corresponding to the F-statistic of one-way ANOVA was lower than 0.05, suggesting that the one or more treatments were significantly different (see Table 2). This was demonstrated by the statistical difference in the DSC comparison between deep learning methods (see Tables 2, 3 where three different multiple comparison correction techniques were used). Nevertheless, the computational cost highlighted that the ENet is much faster than UNet; Table 4 shows the comparison of computational complexity and performance for each model.

Table 2.

ANOVA on the DSC showed statistical differences between segmentation methods

ANOVA	F value	F critic value	P value
ENet versus ERFNet versus UNet	3.667	3.038	0.027
ENet versus ERFNet	5.520	3.907	0.020
ERFNet versus UNet	4.474	3.907	0.036
ENet versus UNet	1.270	3.907	0.261

Table 3.

Tukey HSD, Scheffé and Bonferroni/Holm tests (post hoc tests) were used as multiple comparison correction techniques

Tukey HSD	Q-statistic	P value
ENet versus ERFNet	2.451	0.019
ERFNet versus UNet	2.3413	0.022
ENet versus UNet	0.109	0.189

	T-statistic	P value
Scheffé
ENet versus ERFNet	1.733	0.022
ERFNet versus UNet	1.656	0.026
ENet versus UNet	0.077	0.199
Bonferroni/Holm
ENet versus ERFNet	1.733	0.025
ERFNet versus UNet	1.656	0.030
ENet versus UNet	0.077	0.282

Table 4.

Comparison of computational complexity and performance of the three models

Model name	Number of parameters		Size on disk (MB)	Inference times (s)
	Trainable	Non-trainable		CPU	GPU
Enet	362,992	8352	5.8	122.56	15.23
ERFNet	2,056,440	0	25.3	157.53	16.64
Unet	5,403,874	0	65.0	1398.23	39.11

Figure 4 shows the profiles of training DSC and Tversky loss function for one fold. Both DSC and Tversky loss profiles indicate that the ENet model converges much faster than both the ERFNet and UNet, with the ENet model reaching a training DSC of nearly 0.9 in less than 15 epochs. This suggests that training could be obtained faster with ENet. The fact that the training loss of UNet model is lower than both the ENet and ERFNet suggests the presence of overfitting, although the number of UNet filters in each layer was reduced.

Fig. 4.

Plot the training DSC and loss function Tversky loss for each of three models for one particular fold

It is clear that the utilize of 5 × 5 filters in our UNet architecture implementation is more memory demanding than the 3 × 3 convolution filter in the original UNet. In terms of DSC, we observed, on the first training/testing fold, that the difference on the utilize of 3 × 3 filters versus the 5 × 5 filter were 92.09 ± 3.05% and 93.26 ± 2.53%, respectively. Although the results are close, the 5 × 5 version produces better results with less variance. In terms of computational complexity, the 3 × 3 version is obviously smaller. However, even with 3 × 3 filters, the UNet model has 1,946,338 total trainable parameters compared to ENet with only 362,992 parameters. That’s still a fairly significant difference with a factor greater than 5 times.

Figure 5 displays contours of automatically segmented aorta for different height of the CT axial plane of the aneurysmal aorta. Differences in the capability of all models to segment the whole aorta can be observed. For one representative patient case, Fig. 6 highlights segmented dilated aorta and its valve at fully opened shape as obtained by each deep learning models. Implemented algorithms segmented the whole CT sequence, while the manually reference standard stopped earlier. For this reason, the distal ends of the segmented vessels were wrong due to lack of standard references. Nonetheless, all deep learning models were able to capture the shape of the aortic valve at fully-opened shape. The regions of pronounced curvature changes exhibited geometrical changes between predicted and manually-segmented ATAAs.

Fig. 5.

Comparison of segmentation performance for the three architectures in 8 different slices. The manual segmentation (yellow), ENet (red), ERFNet (blu) and U-Net (green) are superimposed. (Color figure online)

Fig. 6.

Comparison of 3D segmentation of prostate using the three Net architectures. The manual segmentation (yellow), ENet (red), ERFNet (blu) and U-Net (black) are superimposed. (Color figure online)

Discussion

In this study, the feasibility and efficacy of three deep learning models for the segmentation of the aneurysmal ascending aorta was assessed accounting for accurate vessel segmentation, fast training time, low hardware requirements for inference, and low training data requirements. Using ECG-gated CT angiography of 72 patients with ATAA and different valve morphologies, all deep learning models were able to accurately segment the dilated aortas when compared to those obtained by manual segmentation. Among tested deep learning models, the key differences are 1) the ENet and UNet result more accurate than ERFNet, with the ENet faster than UNet; 2) the ENet model converges faster than both the ERFNet and UNet. Although validation in large image dataset is warrant, the clinical application of deep learning may revolutionize the way we diagnose the aneurysmal ascending aorta and open the way towards clinical decision-support system for risk stratification and patient management.

A deep learning model for multiple structures would be consistent and fast to reproduce the same result every time. It is recognized that certain segmentation outputs of our ATAA models revealed lower DSC, but on further inspection, this was due to ground truth annotation error by human readers or artifacts caused by the high heart rate that is common in the bicuspid patient population [2]. Furthermore, the agreement between deep learning predictions and manual segmentations is comparable to that usually reported for the inter- and intra-reader agreements by manual operators [24]. To the best of our knowledge, this is the first study that adopted ENet and ERFNet for the cardiovascular medical imaging analysis. These models are developed for real-time applications and are therefore smaller and faster than the UNet model used in other studies for cardiovascular segmentation [8, 9, 25, 26]. The ENet model has an order of magnitude fewer parameters than both ERFNet and UNet while ERFNet has less than half the number of parameters compared to UNet. Using a fair GPU hardware, we found that the ENet needs only 15.2 s for the ATAA segmentation as compared to the slower 39.1 s shown by the UNet. However, when computations are performed on CPU, the size of the ATAA model has a remarkable impact on the performance of deep learning model, with ENet and UNet, which respectively employ on average about 122.5 s and 1398.2 s to segment the CT data set of a patient with ATAA.

The fact that the proposed deep learning models were able to accurately segment the dilated aortas on the basis of a small training dataset is due to an ad-hoc pre-processing that was previously developed by our group [22, 23, 27, 28]. In general, deep learning methods suffer when applied to class imbalanced data and tends to predict most voxels as belonging to a background class. To overcome this issue, we used a custom version of Tversky loss function [20] to provide a larger weight to the target voxels and thus to learn the foreground object representation more effectively. Moreover, a five-fold cross-validation strategy using 2D slices from all patient cases as model input was used to overcome the limit of the small training dataset while the overfitting was reduced by six different types of data augmentation techniques. In other deep learning studies, the number of data for training was remarkable [6, 8, 9, 26]. With regards to cardiovascular anatomies, Baskaran et al. [9] applied a UNet-inspired deep learning model to segment cardiac structures and great vessels from 206 patients who underwent coronary CT angiography. They obtained an overall median DSC of 0.820. For the abdominal aorta, Roth et al. [26] trained a deep learning model using 331 CT scans to obtain a DSC of 0.79. Another study based on convolutional neural networks segmented three parts of the thoracic aorta, with DSC ranging from 0.83 to 0.88 [25]. Using an ad-hoc pre-processing strategy, we were able not only to reduce the need for a big training dataset but also improved the segmentation accuracy as the DSC was greater than 0.88 with all deep learning models.

The clinical decision-making process for the management of patients with ATAAs is based on the maximum aortic size normalized by the patient body size index or height. However, size is not a good predictor of aortic rupture or dissection [4]. Phenotypic classification has evinced that ATAAs confined to the aortic root grow differently by aneurysm shapes involving the tubular portion of the ascending aorta, thereby demonstrating the lack of predictive capability of the single aortic size measurement [29]. There is therefore an emerging interest in the development of image-derived strategies to improve ATAA risk definition to highly individualized level [30]. These novel strategies rely on flow analysis computed by in vivo 4D Flow MRI [31], computational predictions based on rupture potential indices of the ATAA wall [32], combination of computational analyses and plasma-based biomarkers [15, 33, 34]. Recently, few research groups have proposed machine learning and statistical shape analysis to investigate the relationship between shape features and numerically predicted risk variables of ATAAs [35, 36]. On the other hand, aortic strain for stiffness-based risk predictions is increasing the interest of many researchers because this metric can be easily obtained by echocardiographic imaging without the need of in silico simulations and assumptions on ATAA material properties. We recently developed a mathematical algorithm to quantify the full-field aortic strain of the ATAA wall from ECG-gated CT angiography and predicted the aneurysm risk by a stiffness-based parameter [37]. This approach was time-consuming because it was based on manual segmentations of the ATAA wall at ten cardiac phases. The combination of the proposed deep learning models with mathematical algorithms for strain analysis can be easily implemented in the clinical framework to provide stiffness-based risk prediction and tailor personalized approach to ATAA management. Therefore, this study adds another brick towards the implementation of fully-automatic risk strategies for patients with ATAAs.

There are a number of limitations in this work. The number of patients used for training and validation may have limited the accuracy of deep learning models. The different patterns of aortic dilatation (i.e., aortic root vs. tubular aortic dilatations) and bicuspid phenotypes (i.e., anterior vs. posterior) may have increased the variability in the investigated CT image dataset. As more patients will be recruited, the training and validation of deep learning models will be re-evaluated by grouping patients according to similar shape features or aortic valve phenotypes. In this study, 3D segmentation was not adopted as this approach requires larger dataset and is memory-demanding. One approach to deal with 3D segmentation drawbacks is to down-sample the data or adapt 3D integration of 2D convolutional neural networks trained on orthogonal planes to provide a final 3D segmentation. As this study focused on the impact of different deep learning methods on the segmentation accuracy of dilated aortas, the efficacy and accuracy of 3D approach will be investigated in future studies. Finally, the accuracy of trained deep learning models is likely confined to our CT scanner, and extension to healthy non-aneurysmal aortas could be not straightforward.

Conclusion

This study demonstrated the feasibility and efficacy of deep learning for the segmentation of ATAAs as collected from ECG-gated CT angiography. The tested deep learning models highlighted a good segmentation accuracy with DSC of 88% in all models (i.e., UNet, ENet and ERFNet), with differences related to the training time and data requirements. The clinical application of deep learning for automatic vessel segmentation can improve not only the diagnosis of ATAAs but can also improve the management of patients towards personalized risk strategies.

Compliance with ethical standards

Conflict of interest

Albert Comelli declares that he has no conflict of interest. Navdeep Dahiya declares that he has no conflict of interest. Alessandro Stefano declares that he has no conflict of interest. Viviana Benfante declares that she has no conflict of interest. Giovanni Gentile declares that he has no conflict of interest. Valentina Agnese declares that she has no conflict of interest. Giuseppe M Raffa declares that he has no conflict of interest. Michele Pilato declares that he has no conflict of interest. Anthony Yezzi declares that he has no conflict of interest. Giovanni Petrucci declares that he has no conflict of interest. Salvatore Pasta declares that he has no conflict of interest.

Ethical statement

All procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975, as revised in 2000 (5). Informed consent was obtained from all patients for being included in the study. Additional informed consent was obtained from all patients for which identifying information is included in this article.