Judy Shum1, Satish Muluk2, Adam J. Doyle3, Ankur Chandra3, Mark Eskandari4, Ali Azizzadeh5, Ender A. Finol6
1Carnegie Mellon University, Pittsburgh, PA; 2West Penn Allegheny Health System, Pittsburgh, PA; 3University of Rochester Medical Center, Rochester, NY; 4Northwestern Memorial Hospital, Chicago, IL; 5The University of Texas Health Science Center at Houston, Houston, TX; 6Biomedical Engineering, The University of Texas at San Antonio, San Antonio, TX.
OBJECTIVES: To evaluate a set of global geometric features describing the size, shape and wall thickness of abdominal aortic aneurysms (AAA), in order to develop a learning algorithm model capable of discriminating between ruptured (rAAA) and unruptured AAA (uAAA).
METHODS: An in-house CT image segmentation and geometry quantification code was developed to automatically and non-invasively compute 77 features that quantify the geometry of 90 unruptured (Dmax=5.3±1.2 cm) and 90 ruptured (Dmax=7.0±1.9 cm) AAA. Statistical models were generated to determine patterns and relationships in the data that may not be apparent to the human eye, to yield an optimized model that can classify new data and provide the features that discriminate between different uAAA and rAAA.
RESULTS: The best models were obtained using a Support Vector Machine (SVM) with Chi Square feature selection or k Nearest Neighbors (kNN) with no feature selection and either 4 or 6 neighbors. Though both the SVM and kNN correctly classified 82.2% of the datasets, the ratio of correctly classified rAAA and uAAA was different. SVM correctly classified more uAAA (81% vs. 78% for kNN) while kNN classified more rAAA (87% vs. 83% for SVM). Using the standard maximum diameter of 5.5 cm, 31.1% of the AAAs would have been misclassified (20 ruptured and 36 unruptured). The ruptured and unruptured AAA features tended to group together, with the most predictive features forming clusters. The ten most predictive features, given by the χ2 test, were: skewness of Gaussian curvature distribution, tortuosity, non-fusiform index, minimum Mean curvature, L2 norm of Mean curvature, wall surface area, L2 norm of Gaussian curvature, isoperimetric ratio, length, and the maximum diameter.
CONCLUSIONS: SVM and kNN classified more AAA correctly than a maximum diameter of 5.5 cm, indicating that maximum diameter alone is not sufficient for predicting AAA rupture risk compared to a robust geometry quantification approach.
AUTHOR DISCLOSURES: A. Azizzadeh, Nothing to disclose; A. Chandra, Nothing to disclose; A. J. Doyle, Nothing to disclose; M. Eskandari, Nothing to disclose; E. A. Finol, Nothing to disclose; S. Muluk, Nothing to disclose; J. Shum, Nothing to disclose.
Posted April 2012