Abstracts

RETURN TO ABSTRACT LISTING


1346. Preoperative Classification of Soft-Tissue Neoplasms Using Machine-Learning Augmented Radiomics-Based Decision Classifiers
Authors * Denotes Presenting Author
  1. Brandon Fields *; Keck School of Medicine of University of Southern California
  2. Natalie Demirjian; Keck School of Medicine of University of Southern California
  3. Darryl Hwang; Keck School of Medicine of University of Southern California
  4. Bino Varghese; Keck School of Medicine of University of Southern California
  5. Steven Cen; Keck School of Medicine of University of Southern California
  6. Vinay Duddalwar; Keck School of Medicine of University of Southern California
  7. George Matcuk; Cedars-Sinai Medical Center
Objective:
Radiomics analysis has shown promise in a variety of prognostic applications, including evaluation of response assessment and prediction of relapse. Focused approaches for differentiating between select tumor subgroups have been preliminarily established; however, few attempts have been made at the development of a generalizable radiomics-based prediction model using data pooled from multiple study centers. Our aim was to evaluate the role of using MRI-based radiomics models in differentiating between benign and malignant soft-tissue neoplasms.

Materials and Methods:
128 histologically diagnosed benign (n=36) and malignant (n=92) soft tissue masses were retrospectively identified. Patients were selected as having an available MRI prior to any resection or therapy. Well-differentiated liposarcomas and atypical lipomatous neoplasms were excluded, as these tumors are known to share many histologic and MRI imaging features. No other tumor histologies were excluded. The most common malignant tumors were undifferentiated pleomorphic sarcoma (n=29) and myxoid liposarcoma (n=15), while the most common benign tumors were aggressive fibromatosis (n=10) and schwannoma (n=9). MRIs were uploaded to a dedicated Synapse 3D workstation for manual segmentation under the supervision of a subspecialty-trained musculoskeletal radiologist. Automated data processing scripts aided in parametric extraction of whole-voxel data corresponding to the 3D regions-of-interest across multiple co-registered sequences. Our institutionally developed Radiomics Pipeline was then used to extract 1708 radiomics metrics derived from 9 different texture methods. These features were utilized to train and test a machine-learning algorithm capable of differentiating between benign and malignant lesions. Random Forest (RF) and Real Adaptive Boosting (Adaboost) with a 10-fold cross-validation were employed to establish the prediction models. Area-Under-The-Curve (AUC) was used to quantify the prediction accuracies.

Results:
RF and Adaboost showed very similar performance in differentiating between benign and malignant soft tissue tumors, with AUCs of 0.72 (95% CI 0.63–0.81) and 0.77 (95% CI 0.68–0.85), respectively.

Conclusion:
Our results suggest that machine-learning augmented models can accurately classify unknown soft-tissue neoplasms by identifying potentially high-grade features of the radiomics texture. A machine-learning augmented approach to screening soft tissue masses could help in optimizing treatment strategies for benign masses that may be safe for surveillance, as opposed to those that would benefit from more aggressive management.