Background:
Aortic aneurysm and dissection (AAD) are highly lethal conditions involving progressive weakening, dilation, and tearing of the aortic wall, often resulting in fatal rupture. Currently, AAD is diagnosed solely by measuring vessel diameter using non-invasive imaging techniques, such as ultrasound, computed tomography (CT) angiography and magnetic resonance angiography. Although these techniques exhibit very high sensitivity for the detection of aortic enlargement, they are unable to detect aortic degeneration that occurs before changes in diameter are evident. Early detection of aortic degeneration and prediction of disease progression are critical for improving outcomes in these patients. We have previously shown that contrast-enhanced computed tomography (CECT) using a liposomal-iodine nanoparticle contrast agent (nCECT), enabled early detection of aortic degeneration before aortic enlargement and gross changes are evident. Specifically, the nanoparticle CT contrast agent (NPCA) accumulated in aortic wall at regions of compromised aortic barrier integrity (Figure 1 A-B). In this work, we investigated whether radiomic analysis of nCECT images, referred to as CT nano-radiomics, and machine learning could reveal AAD progression leading to aortic rupture.

Hypothesis/Objective:
To test the hypothesis that patterns of intramural NPCA accumulation could serve as an in vivo biomarker for predicting imminent aortic rupture.

Methods:
C57BL/6 mice (n=44) were challenged with a high-fat diet for 8 weeks and subcutaneous angiotensin-II (Ang II; 2000 ng/min/kg) infusion during the last 1 week. At the end of the challenge, animals were placed on a normal diet and monitored for sudden death from aortic rupture for 28 days. Mice were intravenously administered NPCA (1200 mg I/kg) one day after the start of the Ang II challenge and underwent in vivo nCECT at 3 days post-contrast delivery (Figure 1C). Aortas were segmented on nCECT, and radiomic features (RFs) were calculated using the PyRadiomics Python package [1]. Furthermore, engineered RFs were extracted from Radial Profile Volumes (RPVs) calculated as a stack of 2D radial intensity profiles from the segmented smoothed aorta, starting from the root and extending along the aortic centerline. Intramural NPCA contrast patterns within the tubular wall could be analyzed more accurately by transitioning from standard Cartesian coordinates to RPVs. The analysis utilized only Image Biomarker Standardization Initiative (IBSI) recommended RFs [2]; cross-correlation was employed to eliminate highly correlated RFs. Subsequently, an ANOVA-based feature-importance technique was used for selecting features. Machine learning (ML) was performed on 65% of the data for training and tested on the remaining 35%.

Results:
Nanoparticle contrast-enhanced CT (nCECT) revealed NPCA accumulation within the tubular aortic wall in all subjects. Of the 44 mice included in the study, 11 mice died prematurely due to aortic rupture, while 32 mice survived the four-week study; CT data from one mouse was excluded due to data corruption (Figure 1 D). Two machine learning models were developed to predict these outcomes: a binary classifier for overall survival and a time-to-event survival analysis. A set of 33 models was trained using 3-fold cross-validation. Starting with 44 uncorrelated RFs, we applied repetitive ANOVA and a 1:3 cost function to address class imbalance. Feature elimination reduced the set to the top 14 RFs (6 original, 8 RPV-derived) without loss of accuracy. The SVM quadratic classifier performed best, identifying RPV original GLSZM Low Grey Level Zone Emphasis, Original RV Shape Least Axis Length, and RPV original GLCM IMC2 as the most significant predictors. This model achieved a test set AUC of 0.93, with 80.0% accuracy and an F1 score of 79.8% (Figure 1 right panel).

Conclusion:
Integration of machine learning with engineered radiomic analysis of NPCA images, specifically utilizing RPV, resulted in a highly effective approach for predicting the imminent risk of aortic rupture in a murine model of AAD.

Image/Figure:
Click to view full size
Click to view full size

Image/Figure Caption:
Figure 1. Left: (A-B) Coronal nCECT of thoracic region; green arrows indicate the location of intramural wall CT signal due to intramural NPCA accumulation. A) shows a non-survivor animal, B) shows an animal that survived until the end of the study, C) study plan, D) table listing the number of animals and time of death. Right: Table for training and test performance matrices for the best classifier.

Author
Zbigniew Starosolski
Texas Children’s Hospital