Room: Exhibit Hall | Forum 9
Purpose: To identify radiomic texture features of musculoskeletal masses and develop a machine learning-based predictive model to classify musculoskeletal tumors and compare it to expert radiologist reads.
Methods: MR images from a patient cohort with histologically verified musculoskeletal masses (34) were analyzed and used to train predictive classification models. Additional patients (45) with blinded pathology results were used as an independent validation cohort. Texture and geometrical features were extracted from segmented lesions using 13 directional gray level co-occurrence matrices (GLCM), 256 gray level quantization, and 3 pixel offsets. 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). Expert radiologists independently read the exams and comparison to their accuracy was performed. Receiver Operator Characteristic (ROC) metrics, and area under the curve (AUC) was used to assess predictive accuracy of selected features and classification scheme.
Results: Feature selection substantially reduced the dimensionality of the features for predictive model training from 826 to fewer than 10 prominent features. The highest performing models in terms of AUC were the SGD wrapper+MLP classifier and the SGD wrapper+SGD classifier with 94.8Â±1.1% and 94.1Â±2.4% accuracy, respectively. The results of the additional 45 patient testing data set showed that all models had a classification accuracy of â‰¥87% (39/45), with the SGD wrapper+SGD classifier being the top performer, exhibiting an accuracy of 96% (43/45). Radiomics performed superior to expert radiologists reads.
Conclusion: Radiomic predictive models show promise in characterization of musculoskeletal tumors and machine learning performs better than expert radiologists reads