Program Information
Organ-At-Risks Gross Error Detection Using Classification-Based Anomaly Detection Techniques
H Nourzadeh1*, C Hui1 , W Watkins1 , M Ahmed1 , N Sadeghzadehyazdi2, J Siebers1 , (1) University of Virginia Health Systems, Charlottesville, VA, (2) University of Virginia, ECE Department, Charlottesville, Virginia
Presentations
SU-K-605-8 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: 605
Purpose: To investigate the performance of classification-based techniques for the purpose of gross error detection in delineated organs-at-risk prior to their use in radiation treatment plans, radiomics studies, or bulk big data analysis.
Methods: From 296 locally advanced lung and prostate patients, shape and image related features were computed for 2171 normal manually delineated OARs of 18 different types. Shape features were comprised of estimates of 3D orientation and eccentricity, volume, surface area, relative centroid position, relative position and volume with respect to the external contour. Image features included largest, smallest, and average densities, as well as histogram of oriented gradients (HOG) that captures the orientation of intensity gradient and serves as a localized shape descriptor. To circumvent overfitting, dimensionality reduction was performed on HOG using principal component analysis (PCA) retaining 98% of variance. Three multi-class classifiers, support vector machine (SVM), discriminant subspace ensemble (DSE), and artificial neural network (ANN), were trained and tested using a ten-fold cross-validation scheme.
Results: The accuracy of the classifiers was 96.2%, 99.68%, and 100% for SVM, DSE and ANN, respectively. Imperfect F1-score was observed in 10 and 3 OARs for SVM, DSE classifiers. OARs with gross errors reflected in the features take on low confidence scores, and were reported for review.
Conclusion: The selected features distinctively characterize the OARs involved in this study. Anomalous delineated OAR with common gross errors including mislabeling, undesirable isolated regions, missing contours, and other gross errors in the ROI’s geometry, position and orientation detected using classification-based anomaly detection methods. The performance of DSE and ANN were superior compared to SVM.
Contact Email: