MRI Radiomics-based Machine Learning to Predict Neoadjuvant Chemotherapy Response in Ewing's Sarcoma: Preliminary Results

Purpose or Learning Objective: To evaluate two-dimensional (2D) and three-dimensional (3D) magnetic resonance imaging (MRI) radiomics-based machine learning prediction of neoadjuvant chemotherapy response in Ewing's sarcoma.

Methods or Background: A total of 30 patients were included retrospectively at two tertiary bone sarcoma centers. Inclusion criteria were (1) biopsy-proven Ewing's sarcoma treated with neoadjuvant chemotherapy before surgery; (2) preoperative MRI available; and (3) therapy response evaluated after surgery based on pathologic findings. Seven patients were poor responders; 23 were good responders. On T1-weighted and T2-weighted images, manual segmentations were performed by drawing both 2D regions of interest (ROIs) along tumor borders on the slice showing the largest diameter and 3D ROIs including the whole volume. A total of 1,702 3D and 958 2D features were extracted. Feature stability was assessed through small geometric transformations of the ROIs mimicking multiple manual delineations, and an intraclass correlation coefficient > 0.75 defined feature stability. Feature selection included collinearity and significance analysis. Three machine learning classifiers were considered, such as k-nearest neighbors (k-NN), logistic regression (LR), and random forest (RF). To evaluate the unbiased performance of the classifiers, a cross-validation approach was used with a holdout partition of 80 to 20 (80% for training and 20% for test, repeated 100 times). Class balancing was performed to oversample the minority (i.e., poor responders) class in the training cohort.

Results or Findings: A total of 1,303 3D and 620 2D radiomic features were stable to geometric transformation of the ROI. Four 2D and four 3D features were selected during dimensionality reduction. LR built on 3D features achieved the best performance with 85% sensitivity, 87% specificity, and 85% accuracy (area under the curve [AUC]: 0.9) in predicting response to chemotherapy.

Conclusion: Machine learning showed very good performance in predicting the response of Ewing's sarcoma to neoadjuvant chemotherapy using MRI radiomic features.

Publikationsverlauf

Artikel online veröffentlicht:
02. Juni 2022