Table 2. Radiomics studies on aortic aneurysm and dissection.
| Study | Objective | Study population | Algorithm | Results | Comment |
|---|---|---|---|---|---|
| Monti et al. (7) | To evaluate the CNN algorithm performance in aortic diameters measurement in a heterogeneous population |
250 ECG-gated CT +/− contrast data from heterogeneous population with or without aortic pathologic findings | Manual and AI using CNN measurement of aortic diameters at 9 locations and maximum aortic diameter were made | No significant difference was found in maximum aortic diameter between manual and automatic measurements (P=0.48) | Strength: good reproducibility at 9 AHA sites and maximum aortic diameters were demonstrated using CNN algorithm |
| Weakness: the reproducibility was affected by contrast enhancement, site, positioning, and local pathologic finding. There was a tendency in measurement of the internal diameter by the automatic software. Presence of aortic dissection makes it difficult to detect and measure the aortic diameter because of the presence of two lumina | |||||
| Wobben et al. (8) | 3 segmentation models using 3D residual U-Net for segmentation of the TL, FL, and FLT | Retrospective study of 164 CTA scans from 43 patients with uTBAD who underwent baseline and surveillance imaging | Segmentation models using 3D residual U-Net | DSC of 0.85 (0.77–0.88) and 0.84 (0.82–0.87), for TL and FL respectively | Strength: automated and robust segmentation of the TL and FL |
| Weakness: small sample size; difficulty to differentiate between slow blood flow and FLT or a thick flap and thrombus around the flap causing the reduced reproducibility of FLT segmentation | |||||
| Guo et al. (9) | Clinical data-driven machine learning was used to predict the in-hospital mortality of acute aortic dissection using AI | A cohort of 1,344 acute aortic dissection patients (1,071 patients in the survivor group and 273 patients in non-survivor group) | ML algorithms using logistic regression, decision tree, K nearest neighbor, Gaussian naive bayes, and XGBoost were included | XGBoost model was the most effective model with the greatest ROC (0.927, 95% CI: 0.860–0.968) | Strength: 5 ML models were trained and developed to predict the in-hospital mortality. Comparison was made among the 5 ML models |
| Weakness: small dataset collected from a single source | |||||
| Lyu et al. (10) | Deep learning-based algorithm to segment dissected aorta on CTA images | The CTA volumes for training and evaluation were collected from 42 AD patients including 37 males and 5 females | 3D CNN is applied to divide the 3D volume into two anatomical portions | Dice index >92% on average | Strength: the segmentation performance improved using combination of 3D and 2D models Weakness: variations in the segmented arch parts led to notable errors on some CTA volumes; lower accuracy due to multiple local intimal flaps |
| Adam et al. (11) | Evaluate an automatic, deep learning-based method ARVA to obtain maximum aortic diameter, to give cross sectional outer to outer aortic wall measurements | Training database consist of 489 CTA which was used to train for automatic external aortic wall measurements using neural networks. Another database of 62 CTA was used for validation | Deep learning-based method ARVA | The median absolute difference with reference to expert’s measurements ranged 1–2 mm among all annotators, while ARVA reported a median absolute difference of 1.2 mm | Strength: ARVA measurement on the diameter in the plane perpendicular to the centerline and automatic method in maximum aortic diameter fall within the inter-annotator variability |
| Weakness: Small samples size with only 62 scans involved; performances were reduced on portal phase scanners; the algorithm can output outlier measurements that are out of the variability range of annotators | |||||
| Beetz et al. (12) | AI-based tissue segmentation to analyze body composition using in patients with Marfan syndrome | 25 patients aged ≤50 years with Marfan syndrome without prior aortic repair | There was a significantly increased SMD measured in HUs in patients with aortic enlargement using AI-based automated analysis of body composition at L3 | A significant prediction of aortic enlargement for SMD (P=0.05) and PMI (P=0.04) was shown using multivariate logistic regression | Strength: CT angiography using AI-based analysis of body composition at L3 was feasible and easily available in Marfan patients |
| Weakness: retrospective dataset and the relatively small cohort; muscle density on CT reflecting the body fitness and muscles structure may be influenced by the amount of contrast medium uptake | |||||
| Zhou et al. (13) | Investigation on the diagnostic value of CT scan-based radiomics model for acute aortic dissection | 50 patients with acute aortic dissection +50 non-dissection patients were retrospectively selected (70 in the training group and 30 in the validation group) | 3 types of characteristics (first-order statistics, geometric descriptive features, and texture features) were extracted using PyRadiomic | Accuracy, sensitivity, specificity, and AUC of the training group were 94.3%, 91.2%, 97.2%, and 0.988 (95% CI: 0.970–0.998), respectively. The respective values for the validation group were 90.0%, 94.1%, 84.6%, and 0.952 (95% CI: 0.883–0.986) | Strength: non-contrast-enhanced CT imaging is convenient; good performance and high accuracy |
| Weakness: only the largest lesion layer was used, thus unable to cover all the layers in the lesion in the ROI | |||||
| Bratt et al. (14) | Evaluate the performance of trained standard deep learning segmentation model in measurement of aortic volume and diameter | Deep learning training set consisted of 3,051 CTA volume were used. Quality evaluation on 50 scans and 57 with 207 scans patients was used to evaluate the temporal reproducibility | The 2D deep learning segmentation model using U-Net backbone with an EfficientNet-B3 encoder of depth seven | Deep learning temporal showed better reproducibility for measures of both volume (P<0.008) and diameter (P<1e-5) compared with previously reported values of manual inter-rater variability | Strength: deep learning temporal showed better reproducibility for measurement |
| Weakness: manual plane selection was required to extract measurements; not all clinically relevant and canonical aortic stations diameters were assessed | |||||
| Yu et al. (15) | Segmentation and diameter measurement of TBAD using automatic method | 139 consecutive patients with TBAD CT angiographic images | 3D deep CNN network | Mean dice coefficient scores were 0.958, 0.961, and 0.932 for EA, TL, and FL, respectively | Strength: DL method showed better performance with higher accuracy for the EA, TL, and the FL at all positions compared with the manual method |
| Weakness: only TBAD patients involved; not included TBAD with thrombus in the FL | |||||
| Harris et al. (16) | Assess the likelihood of aortic dissection and rupture in prospective patients using CNN trained model | 34,689 post-contrast chest CT data was studied using the aortic dissection model | CNN model trained on aortic dissection and rupture | Sensitivity and specificity of 87.8% and 96.0% for aortic dissection and 100% and 96.0% for aortic rupture | Strength: high-volume inference system for identifying and prioritizing aortic dissection and rupture studies was implemented |
| Weakness: some abdominal aortic dissections in abdomen-pelvis studies intentionally not routed through the aortic injury model might have missed; difficult to identify on post-contrast CT images where dissections near the aortic valve ended and where the valve itself began because both can look similar | |||||
| Wojnarski et al. (17) | Using ML to identify the patterns of aortopathy and establish the association with valve morphology and patient characteristics | 3D CT reconstruction for 656 patients undergoing ascending aorta surgery with BAV | To cluster aortic dimension using unsupervised partitioning around medoids. Polytomous random forest analysis was used to identify the group differences | Phenotypes were visually recognizable with 94% accuracy | Strength: unsupervised analysis has yielded aortic phenotype descriptors that are sufficiently distinct to be reliably recognized by rather simple visual rules |
| Weakness: patients in the studies had advanced aortopathy requiring surgical ascending aorta replacement, and might not be representative of all patients with a bicuspid aortic valve | |||||
| Liang et al. (18) | To study the geometric features of the aorta that associated with risk of rupture | A total of 729 shapes were obtained from the shape distribution | SSM | An average risk classification accuracy of 95.58% and an average regression error of 0.0332 | Strength: a novel study on geometric risk stratification derived from the association between shape features and aneurysm risk predictions using FEA with machine learning algorithm |
| Weakness: only using shape features as risk predictors and the other variables were kept constant; the threshold of tissue failure was different among different patients. Difficult to model material strength | |||||
| Guo et al. (19) | To create an effective screening system for thoracic aortic dissection using non-contrast CT-based radiomic features | 378 patients from 4 medical centers underwent non-contrast chest CT scans as well as CT angiography or MRI |
mRMR and the LASSO were used to select the radiomic features | The AUCs was 0.91 (95% CI: 0.86–0.95), 0.92 (95% CI: 0.86–0.98) and 0.90 (95% CI: 0.82–0.98) in the training set, validation set and in the external test set respectively | Strength: the study illustrated the diagnosis of thoracic AD can be effectively derived from non-contrast CT using radiomic signature |
| Weakness: potential selection bias in patients who underwent imaging studies may lead to a high positive rate of thoracic AD |
2D, two-dimensional; 3D, three-dimensional; AD, aortic dissection; AHA, American Heart Association; AI, artificial intelligence; ARVA, augmented radiology for vascular aneurysm; AUC, area under the curve; BAV, bicuspid aortic valve; CI, confidence interval; CNN, convolutional neural network; CT, computed tomography; CTA, computed tomography angiogram; DSC, Dice similarity coefficients; EA, entire aorta; ECG, electrocardiogram; FEA, finite element analysis; FL, false lumen; FLT, false lumen thrombosis; HUs, Hounsfield units; LASSO, least absolute shrinkage and selection operator; ML, machine learning; MRI, magnetic resonance imaging; mRMR, max-relevance and min-redundancy; PMI, psoas mass index; ROC, receiver operating characteristic; ROI, region of interest; SMD, skeletal muscle density; SSM, statistical shape model; TBAD, type B aortic dissection; TL, true lumen; uTBAD, uncomplicated type B aortic dissection; XGBoost, extreme gradient boost.