A Triple-Classification Radiomics Model for the Differentiation of Primary Chordoma, Giant Cell Tumor, and Metastatic Tumor of Sacrum Based on T2-Weighted and Contrast-Enhanced T1-Weighted MRI

Abstract

Background: Preoperative differentiation between primary sacral chordoma (SC), sacral giant cell tumor (SGCT), and sacral metastatic tumor (SMT) is important for treatment decisions.

Purpose: To develop and validate a triple-classification radiomics model for the preoperative differentiation of SC, SGCT, and SMT based on T2-weighted fat saturation (T2w FS) and contrast-enhanced T1-weighted (CE T1w) MRI.

Study type: Retrospective.

Population: A total of 120 pathologically confirmed sacral patients (54 SCs, 30 SGCTs, and 36 SMTs) were retrospectively analyzed and divided into a training set (n = 83) and a validation set (n = 37).

Field strength/sequence: The 3.0T axial T2w FS and CE T1w MRI.

Assessment: Morphology, intensity, and texture features were assessed based on Formfactor, Haralick, Gray-level co-occurrence matrix (GLCM), Gray-level run-length matrix (GLRLM), histogram.

Statistical tests: Analysis of variance, least absolute shrinkage and selection operator (LASSO), Pearson correlation, Random Forest (RF), area under the receiver operating characteristic curve (AUC) and accuracy analysis.

Results: The median age of SGCT (33.5, 25.3-45.5) was significantly lower than those of SC (58.0, 48.8-64.3) and SMT (59.0, 46.3-65.5) groups (χ² = 37.6; P < 0.05). No significant difference was found when compared in terms of genders, tumor locations, and tumor sizes of SC, SGCT, and SMT ( ${\chi }_{\text{gender}}^2=3.75,{\chi }_{\text{location}}^2=2.51,{\chi }_{\text{size}}^2=5.77$ ; P1 = 0.15, P2 = 0.29, P3 = 0.06). For the differential value, features extracted from joint T2w FS and CE T1w images outperformed those from T2w FS or CE T1w images alone. Compared with CE T1w images, features derived from T2w FS images yielded higher AUC in both training and validating set. The best performance of radiomics model based on joint T2w FS and CE T1w images reached an AUC of 0.773, an accuracy of 0.711.

Data conclusion: Our 3.0T MRI-based triple-classification radiomics model is feasible to differentiate SC, SGCT, and SMT, which may be applied to improve the precision of preoperative diagnosis in clinical practice.

Level of evidence: 4 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2019;49:752-759.

Keywords: LAVA-Flex; least absolute shrinkage selection operator; machine-learning; radiomics; random forest; sacrum.