E1397. Prediction of Adverse Outcomes After Surgical Excision of Extremity Malignant Soft Tissue Tumors Using Radiomics
  1. Max Hao; Stony Brook University Renaissance School of Medicine
  2. Connor Cowan; Stony Brook University Renaissance School of Medicine
  3. Mutshipay Mpoy; Stony Brook University Renaissance School of Medicine
  4. Sahil Rawal; Stony Brook University Renaissance School of Medicine
  5. Elaine Gould; Stony Brook University Renaissance School of Medicine
  6. Prateek Prasanna; Stony Brook University Renaissance School of Medicine
  7. Daichi Hayashi; Stony Brook University Renaissance School of Medicine
Currently no imaging biomarkers exist to predict response to resection in extremity malignant soft tissue tumors such as sarcomas. Prediction of postoperative likelihood of adverse outcomes such as locoregional tumor recurrence and distant metastasis is helpful for treatment planning. To establish predictive markers for these adverse outcomes in extremity malignant soft tissue tumors, we study subvisual textural radiomic signatures extracted from multiparametric diagnostic MRI.

Materials and Methods:
Our retrospective study included 187 patients with extremity soft tissue tumors who had surgical resection between 2011 and 2021. 62 patients had malignant tumors with preoperative and follow-up MRI. Images and patient outcome data were deidentified for subsequent tumor delineation. T1w, T2w, and postcontrast sequences from preoperative MRI were studied. Tumors were delineated on each MRI sequence by two radiology residents and verified by an attending musculoskeletal radiologist. An adverse outcome is considered when there is tumor recurrence or distant metastasis. 3D radiomic features from the gradient, Haralick, Gabor, Laws Energy, CoLlAGe, and raw feature families were obtained from within each annotated region of interest. First-order statistics such as mean, median, skewness, kurtosis, etc., were then computed. Minimum redundancy maximum relevance (MRMR) was used to select features to train machine learning classifiers for adverse outcome prediction. Models using Random Forest Classification (RFC), Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA) were evaluated in a leave-one-out cross validation (LOOCV) setting. Differences in MRMR selected feature expression between patients with and without adverse events were studied using non-parametric Mann-Whitney U test (significance level p < 0.05).

Five out of 75 radiomic features among the sequence groupings were found to have significant differences in expression levels between the adverse and nonadverse outcome groups. Features derived from the T1w sequence using RFC were most discriminative. Our preliminary results suggest that textural radiomic characteristics appraised on MRI are associated with clinical outcomes for resected soft tissue tumors. Machine learning classifiers can utilize features from multiple MRI sequences to predict patient outcomes in this population.

MRI radiomic features of extremity soft tissue tumors can be used to predict the likelihood of tumor recurrence and metastasis. It can be a powerful tool for treatment planning and surveillance in patients with extremity malignant soft tissue tumors.