Program Information
Robust Radiomic Classification Models Using T2-Weighted MRI Geometrical and Texture Features
A Rodriguez*, S Fisher , M Folkert , A Chhabra , J Wang , UT Southwestern Medical Center, Dallas, TX
Presentations
MO-F-CAMPUS-JT-4 (Monday, July 31, 2017) 4:30 PM - 5:30 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To quantify radiomic texture features of T2 fat-suppressed (T2-FS) MR images of benign and malignant soft tissue musculoskeletal tumors, and to develop a machine learning based predictive model to classify benign and malignant lesions based on texture and geometric features.
Methods: T2-FS MR images from non-uniform acquisition protocols of a 34 patient cohort with histologically verified musculoskeletal tumors (16 malignant and 18 benign) were used to train the classification models. Twenty additional patients with blinded pathology results were used as an independent validation cohort. Tumor contours were drawn manually by an expert radiologist. Texture and geometrical features were extracted from segmented lesions using a combination of 13 directional grey level co-occurrence matrices (GLCM), 256 grey level quantization (Q), and 3 different pixel offsets (PO). Feature selection was performed by both Correlation based Feature Selection (CFS) and wrapper-subset schemes using different machine learning strategies including logistic regression (LG), sequential minimal optimization (SMO), multi-layer perceptron (MLP), and stochastic gradient descent (SGD). Receiver operator characteristic (ROC) derived metrics were used to assess the predictive accuracy for each combination of selected features and classification scheme by averaging results from 10 individual 3-fold cross-validations.
Results: Feature selection greatly reduced dimensionality of features for predictive model training from 826 to 10 or less prominent features, especially for wrapper-subset schemes. The highest performing models in terms of area under the curve (AUC) were the SGD wrapper+MLP classifier and SGD wrapper+SGD classifier with 94.8±2.6% and 94.1±2.4% respectively. When subjected to an independent validation cohort of 20 patients, accuracy of classification was >85% for all top performing models with the top SGD+SGD model performing at 95% accuracy.
Conclusion: Radiomic predictive models show promise in determining histology of musculoskeletal tumors. This method may be a useful tool for radiologists to diagnose these tumors prospectively.
Contact Email: