MRI Texture-Based Models for Predicting Mitotic Index and Risk Classification of Gastrointestinal Stromal Tumors.

Background: Treatment regimens and prognoses of gastrointestinal stromal tumors (GIST) are quite different for tumors in different risk categories. Accurate preoperative grading of tumors is important for avoiding under- or overtreatment.

Purpose: To develop and validate an MRI texture-based model to predict the mitotic index and its risk classification.

Study Type: Retrospective.

Population: Ninety-one patients with histologically-confirmed GIST; 64 patients in a training cohort, and 27 patients in a test cohort.

Field Strength/sequence: T -weighted imaging (T WI), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced three-dimensional volumetric interpolated breath-hold examination (3D-VIBE) at 1.5T.

Assessment: GIST images were manually segmented by two independent radiologists using ITK-SNAP software and MRI features were extracted using Pyradiomics. Two pathologists reviewed the tissue specimens of the tumors to identify the mitotic index and risk classification in consensus.

Statistical Tests: The least absolute shrinkage and selection operator (LASSO) regression method was used to select texture features. A logistic regression model was established based on the radiomic score (radscore), tumor location, and maximum diameter to predict tumor classification and develop a nomogram. Receiver operator characteristic (ROC) curves were used to evaluate the ability of the nomogram to distinguish between two tumors with different risk classifications, and a calibration curve was used to evaluate the consistency between the predicted risk and the actual risk.

Results: The texture signature achieved high efficacy in predicting the mitotic index area under the curve ([AUC], 0.906; 95% confidence interval [CI]: 0.813, 0.961). A nomogram for prediction of the risk classification of GIST, which incorporated this texture signature together with maximum tumor diameter and location, allowed good discrimination in the training cohort (AUC, 0.878; 95% CI: 0.769, 0.960) and the validation cohort (AUC, 0.903; 95% CI: 0.732, 0.922).

Data Conclusion: The texture-based model can be used to predict GIST mitotic index and risk classification preoperatively.

Level Of Evidence: 2.

Technical Efficacy Stage: 3.