Endovascular aneurysm repair (EVAR) for abdominal aortic aneurysm (AAA) has some advantages in the minimal invasion, rapid postoperative recovery, shorter hospital stay, and low mortality and morbidity. However, long-term follow-up results reveal a high risk of postoperative complications and re-intervention, when compared with open surgical repair (OSR). Thus, for patients who undergo EVAR, lifetime follow-up is recommended by guidelines (1–4). However, close follow-up may result in unexpected side effects, such as renal function impairment, radiation exposure, and economic and time costs. Therefore, an individualized follow-up protocol is needed (5–9).

Morphological features have been widely used to evaluate risk after EVAR. The operators first treat patients they do not make an estimation of individual patient risk through these adverse morphological features, such as the short neck, angulated neck, intraluminal thrombosis (ILT), calcification, and iliac artery tortuosity (9–11). Recently, Karthikesalingam Alan used preoperative morphological features and artificial neural network (ANN) to predict endograft complications, and Ali Kordzadeh applied 26 preoperative attributes and ANN for the detection of endoleaks (types I–III, respectively). However, the limited number of patients with different complications may have influenced the prediction performance. Moreover, the above studies did not fully exploit the morphological features and did not compare the differences between different machine learning (ML) algorithms, and the feature extraction process was time-consuming (12, 13).

Convolutional neural networks (CNNs) extract non-linear, entangled, and representative features from massive and high-dimensional data from medical images. Benefiting from its strong feature-learning ability, the deep learning (DL) model has shown human expert-level performance in diagnosis, detection, prognosis, and treatment selection (14–16). Unfortunately, preoperative computed tomography angiography (CTA) DL feature stratification methods for EVAR were still unavailable. Radiomics has been widely used in the study of oncology (17). Texture analysis, as a part of radiomics, has gradually been used in AAA, such as predicting expansion, endotension, regression of endoleaks, and aneurysm progression after EVAR (18–21). Compared with texture analysis, radiomics can obtain more features, which may potentially improve prognostic and predictive accuracy in EVAR (22–24). Recently, Charalambous Stavros used radiomics and support vector machines (SVM) trained on 1-month and 6-month radiomic data after EVAR to predict Type 2 endoleaks (T2ELs) leading sac expansion at 1 year (25). Similarly, at present, preoperative CTA radiomic feature stratification methods for EVAR were unavailable.

In this study, we first determined the significant morphological features and developed a morphological feature model based on different ML algorithms. We then used DL algorithms to make the model development more convenient. Subsequently, we studied the relationship between DL features, radiomic features, and EVAR-related severe adverse events (SAEs), to develop models and compare them with an optimized morphological feature model.

Existing risk score models, such as the Glasgow Aneurysm Score (GAS), Vascular Biochemical and Hematological Outcome Model (VBHOM), the National Surgical Quality Improvement Project (NSQIP), the Vascular Study Group of New England (VSGNE), the Vascular Governance North West model (VGNW), the Medicare model, and the EVAR Risk Assessment (ERA) model, cannot accurately predict complications and re-intervention in the era of EVAR (35–41). Fortunately, when studying the risk factors for model development, we found that morphological features were highly correlated with complications and re-intervention after EVAR. In addition, we also found that conventional statistical models could not fully exploit complex and subtle relationships between features for prediction. These findings were confirmed by Karthikesalingam Alan and Ali Kordzadeh. Karthikesalingam Alan applied ANN and 19 preoperative morphological features to predict endograft complications, and in external validation, the 5-year rates of freedom from aortic complications, limb complications, and mortality in low- and high-risk groups were 95.9 vs. 67.9%, 99.3 vs. 92.0%, and 87.9 vs. 79.3%, respectively (10). Ali Kordzadeh used ANN and 26 preoperative attributes (of which five morphological features) for the detection of endoleaks (types I–III, respectively), and the overall accuracy of the model was >86% (11). To avoid the limited number of patients with different complications that influenced the prediction performance and fully exploited morphological features, the prediction performance between different ML algorithms and model development was easier. A total of 486 patients were used for morphological feature model development, of which 151(31.1%) patients had SAEs. We identified 8 significant morphological features from 32 morphological features and developed a morphological feature model based on different ML algorithms. Our morphological feature model based on SVM had better performance than other ML algorithms, with an AUC of 0.76, an accuracy of 0.76, and an F1 score of 0.82. Later, we proposed the DCNN model for fully automatic segmentation of the ILT in preoperative CTA images. Distinct from traditional thrombus segmentation methods, which have been addressed with intensity-based semiautomatic algorithms combined with shape priors (42, 43). Our DCNN model fully automatically segmented ILT and achieved a mean IOU score of more than 90.78% for segmentation of ILT and AAA AL. The volume difference between the segmentation obtained with the proposed DCNNs and the ground truths was within the experienced human observer variance (P > 0.05). Both the ILT segmentation performance and ILT volume difference were better than the currently available models (44, 45).

Development and comparison of multimodal models showed that the optimized morphological feature model based on AdaBoost had a better predictive performance with an AUC of 0.62, an accuracy of 0.69, and an F1 score of 0.81. The AUC, accuracy, and F1 score decreased when compared with our previous studies (AUC 0.76, accuracy 0.76, and F1 score 0.82). We assumed that the main reason was that, in this section, we performed a 5-fold cross-validation, and the limited amount of data affected the model prediction performance (30). However, it did not prevent the morphological feature model as a useful adjunct tool for predicting outcomes after EAVR under the condition of sufficient data. In addition, the morphological feature model may be more acceptable for clinicians.

Based on DL algorithms and graphics processing units (GPUs) that power their training, the DL model has shown human expert-level performance in prognosis (12, 25). Our proposed DL model (Fold 3) predicts the outcome after EVAR with an AUC of 0.82, an accuracy of 0.85, and an F1 score of 0.89. Through a hierarchical neural network structure, the DL model extracted multilevel features from visual characteristics to abstract mappings that were directly related to SAEs after EVAR (29). The DL model was fast and easy to use. Since our method was an end-to-end pipeline that requires only the manually selected AAA region and image preprocessing, the CTA slice was input to predict the status after EVAR directly without further human input. It was easier than the morphological feature model and radiomic method. Besides, the DL model only required CTA images, without increasing the cost. In addition, CTA was easy to acquire throughout the course of treatment; therefore, this model could be used multiple times.

Radiomic methods used CT images to quantify AAA information at the macroscopic level, and radiomic analysis provides quantitative features to mine high-dimensional information and may build the relationship between AAA images and EVAR-related SAEs (20–22). Reviewing the current literature, few radiomic studies have been carried out in patients with AAA. However, there have been many studies about texture analysis, and Carl W. Kotze used CT texture analysis and 18F-fluorodeoxyglucose positron emission tomography to predict AAA dilatation (16). García G developed a computer-supported endotension detection method based on texture analysis for EVAR (17). García G also evaluated the texture features and classification of AAA endoleak after EVAR, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level dependence matrix (GLDM) were able to distinguish favorable or unfavorable regression with an accuracy of 93.41 ± 0.024%, 90.17 ± 0.077%, and 81.98 ± 0.045%, respectively (18). Ding N applied GLCM, GLRLM, and GLDM for predicting aneurysm expansion after EVAR with accuracy (85.17, 87.23, and 86.09%) and AUC (0.90, 0.86, and 0.83) (19). Recently, Charalambous Stavros used radiomics and the SVM classifier trained on 6-month radiomic features to predict T2ELs leading sac expansion at 1 year with an AUC of 89.3%, and the SVM classifier developed with 6-month radiomic features showed an AUC of 95.5% at 1 year (23). Our radiomic model is based on LR to preoperative prediction outcomes after EVAR with AUC 0.93, accuracy 0.86, and F1 score 0.91. Although, compared with other modal models, the radiomic model had better prediction performance. We suggest that the morphological feature, DL, and radiomic models all can be used to predict outcomes after EVAR.

Our study has some limitations. First, due to the retrospective single-center nature of the study, selective bias could not be avoided. Despite internal validation and testing, it was necessary to estimate these multimodal models at other institutions. Second, the limited data affected the prediction performance for the development of multimodal models. Besides, the relatively small sample size limited the possibility of conducting a subgroup predicting different SAEs after EVAR. Third, the morphological feature model development was still time-consuming under the condition of fully automatically segmenting ILT and computing ILT percentage. The DL model showed effective performance for predicting outcomes after EVAR. A high environment was required for DL model development. At present, radiomics lacks a standardized methodology for radiomic analyses, a universal lexicon to denote features that are semantically equivalent, and lists of feature values alone do not sufficiently capture the details of feature extraction that might nonetheless strongly affect feature values (43, 44). Finally, the combination of morphological features, DL, and radiomic features was unclear. In the future study, we will aim to improve prediction performance by combining different modal models.

Applying morphological features, DL, and radiomic models, we can evaluate the risk of postoperative outcomes after EVAR. Using these models, operators can more accurately estimate individual patient risk after EVAR and identify subgroups of patients who require more intensive follow-up. This might affect both patient selection and surveillance after EVAR, which remained important for prognosis after EVAR.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s.

The studies involving human participants were reviewed and approved by Zhongshan Hospital Fudan University Ethic Committee. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

ZS and WF contributed to the conception and design. MZ, YD, and XL performed the data collection and interpretation. YW, ZZ, and ZS performed ROI segmentation and images preprocessing. YW analyzed the datasets and wrote this manuscript. ZS obtained the funding. All authors read and approved the final manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The handling editor ZL is currently organizing a Research Topic with WF.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.