| Potential of Radiomics and optimized statistical machine learning in grading patients with Glioma |
| Paper ID : 1076-ISCH |
| Authors |
|
Mohamed Nabil Sultan *1, Sherif Yehia1, Magdy Mohamed Khalil2 1Faculty of Science - Helwan University 2School of Applied Health Sciences, Badr University in Cairo (BUC), Badr City and Department of Physics, Faculty of Science, Helwan University, Cairo, Egypt. |
| Abstract |
| Abstract Purpose. The purpose of this work was to develop an optimized classification and predictive models to differentiate between grade II brain glioma (LGG) from Grade III brain glioma (HGG) using various machine learning approaches. Material and Method. A number of 135 brain tumor (68 Grade II and 67 Grade III) MRI imaging series were acquired from two different public datasets TCIA-TCGA and TCIA-LGG. All tumors were manually cropped, preprocessed, and segmented. Then many shape, first order, textural and wavelet-based radiomic features were extracted. Dimensionality reduction was performed using Principal Component Analysis (PCA) and two feature selectors K-best and Percentile selectors were used in feature selection as data preparation before introduction into various machine learning classifiers. To reach the highest performance possible models hyperparameter optimization was performed using grid search cross validation then evaluating models based on task and classification performance. Results The best three performance revealed were Linear Discriminant Analysis (LDA), Support Vector Machine (SVM) and Logistic Regression (LR). LDA came first surpassing all other models with both feature selectors. LDA achieved AUROC of 0.961±0.053, accuracy of 0.911±0.043, sensitivity of 0.957±0.043, and specificity of 0.869±0.105 using percentile selector whereas AUROC of 0.955±0.072, accuracy of 0.912±0.035, sensitivity of 0.927±0.058, and specificity of 0.899±0.077, by using K-best selector. Conclusion. Statistical machine learning and optimization approaches have a significantly high discriminative power and are able to show a superb performance in classification of Grade II versus grade III brain gliomas. |
| Keywords |
| Brain gliomas, radiomics, machine learning, optimization |
| Status: Abstract Accepted (Poster Presentation) |