Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2557 - Pre-Treatment Multi-Sequence MRI Radiomics for Predicting Pathologic Complete Response in Sarcoma Patients Treated by Radiation Therapy
Presenter(s)
Adeniyi Olabumuyi, MBBS - University of Maryland Radiation Oncology, Baltimore, MD
H. Moradmand1, A. A. Olabumuyi2, M. H. Brown2, A. J. Allen2, E. Manuel3, A. C. Nwiloh2, D. Kunaprayoon2, W. F. Regine Jr2, J. K. Molitoris2, and L. Ren4; 1University of Maryland School of Medicine, Baltimore, MD, 2Department of Radiation Oncology, University of Maryland School of Medicine, Baltimore, MD, 3Department of Radiation Oncology, University of Maryland Medical Center, Baltimore, MD, 4University of Maryland, School of Medicine, Radiation Oncology, Baltimore, MD
Purpose/Objective(s):
Pathologic complete response (pCR) after neoadjuvant radiation therapy (RT) is strongly associated with improved outcomes in soft tissue sarcoma. However, patient response to RT remains highly heterogeneous, and reliable methods to predict pCR before treatment are lacking, limiting the ability to individualize therapy. In this study, we investigated pre-treatment MRI-based radiomic features, including tumor subregion analyses, for predicting pCR.Materials/Methods:
We retrospectively studied 72 tumors from 71 patients treated with neoadjuvant RT followed by surgery. Radiomic features were extracted from pre-treatment T1-weighted post-contrast (T1c) and T2-weighted MRI. For each tumor, features were calculated from three regions: the whole tumor, a 3-mm intratumoral rim, and the remaining tumor core. We built separate models using T1c features alone, T2 features alone, and both sequences combined. Tumor size, histologic grade, and ECOG performance status were also used for modeling. We used 3-fold stratified cross-validation. All data processing was done using only the training data within each fold. We removed highly correlated features (|r| > 0.90), imputed missing values using the median, excluded low-variance features, and standardized the remaining variables. Feature selection was performed within each training fold using Elastic Net, Mutual Information, or mRMR. Seven features were selected per fold. These features were used to train logistic regression, random forest, and XGBoost models. Model performance was evaluated using out-of-fold area under the receiver operating characteristic curve (AUC). We estimated 95% confidence intervals using 1000 bootstrap samples.Results:
pCR occurred in 26 tumors (36.1%). The best-performing model used both T1c and T2 features. Whole-tumor T1c features selected by Mutual Information and modeled with random forest achieved an AUC of 0.711 (95% CI: 0.560–0.846). In single-sequence models, T1c performed better than T2. The best T1c-only model achieved an AUC of 0.682 (95% CI: 0.480–0.851). The best T2-only model achieved an AUC of 0.594 (95% CI: 0.417–0.763). Radiomic-only models performed as well as or better than models that included clinical variables. Adding tumor size, grade, and ECOG did not meaningfully improve performance. A small increase was seen in one T2 model (AUC 0.599), but no improvement was observed in T1c or combined models. Overall, T1c-derived features, especially from the rim and whole tumor, were more often present in higher-performing models.Conclusion:
Our study demonstrated the efficacy of using pre-treatment MRI radiomic features with machine learning to predict pCR for soft-tissue sarcoma patients. Models using both T1c and T2 features performed better than single-sequence models. Adding standard clinical variables did not substantially improve prediction.