Informatics and Data Science

Multimodal MRI Machine-Learning Radiomics Framework to Standardize Metastasis-Risk Assessment in Soft-Tissue Sarcoma

Location: Concourse Foyer Kiosk

First Author(s)

PM

Peyman Mirghaderi, MD, MPH

Postdoctoral Scholar
University of Washington
Seattle, Washington, United States

Co-Author(s)

EA

Ehsan Alipour, MD PhD

PhD Graduate
University of Washington, United States
NE

Nasim Eshraghi, MD

Postdoctoral Scholar
University of Washington, United States
JW

Jim Wu, MD

Associate Professor
Brigham and Women's Hospital, United States
PK

Paul Kinahan, PhD

Professor
University of Washington, United States
LW

Lawrence White, MD

Professor
University of Toronto, United States
MC

Majid Chalian, MD

Associate Professor
University of Washington, United States
PT

Peter C Thurlow, MD

Assisstant Professor
University of Washington, United States

Purpose: Risk stratification for metastasis in soft-tissue sarcoma (STS) varies substantially across clinicians and institutions, contributing to heterogeneous surveillance strategies and treatment decisions.

Aim: To develop and evaluate an explainable multimodal machine-learning radiomics framework integrating MRI radiomics, radiologist-derived semantic features, and clinical variables to support standardized metastasis-risk assessment in STS.

Methods/Materials: Patients with primary STS, no baseline metastasis, available pre- and post-neoadjuvant MRI, and ≥12-month follow-up were retrospectively included. Tumors were manually segmented on four MRI sequences. Seventeen semantic imaging features were recorded, and radiomic features were extracted using prespecified PyRadiomics settings. Three supervised machine-learning models were trained using XGBoost: radiomics+clinical (RC), semantic+clinical (SC), and combined radiomics+semantic+clinical (RCS). Performance was evaluated using nested 3-fold cross-validation to minimize optimistic bias. Model discrimination was assessed by AUROC, and SHapley Additive exPlanations (SHAP) were used to evaluate feature contribution consistency and interpretability across folds.

Results: Among 246 patients, 102 (41.5%) developed metastasis. The multimodal RCS model (44 features) performed with an AUROC of 0.82 on the training set and 0.68 on cross-validation. The RC model (6 features) achieved an AUROC of 0.79 on the training set and 0.66 on cross-validation, while the SC model (24 features) had an AUROC of 0.77 on the training set with an average cross-validation score of 0.69.
SHAP analysis demonstrated stable, clinically intuitive drivers of predicted risk across folds, including tumor size, necrosis-related semantics, and selected texture features.

Conclusions: An interpretable machine-learning radiomics framework using prespecified feature extraction and fold-wise explanation provides moderate discrimination while enabling transparent and auditable metastasis-risk assessment. These findings highlight the potential value of explainable AI for standardizing oncologic risk stratification across radiology practices, even when incremental performance gains are limited.