Decision-Tree Classification of Chordomas and Chondrosarcomas

Background: Primary neoplasms of the clivus are uncommon and can be difficult to diagnose radiographically. Differentiation of chordomas and chondrosarcomas has clinical relevance for treatment and prognosis.

Objectives (1) Develop a decision-tree model for radiographic differentiation of chordomas and chondrosarcomas; (2) determine most predictive diagnostic criteria; (3) validate the model.

Methods: Patients with pathologic confirmation of chordoma or chondrosarcoma of the skull base were identified in our tumor registry. Retrospective review of preoperative CT and MRI was performed by neuroradiologists using possible diagnostic criteria posed as yes/no questions. A stratified sample of 70% of records was used as a training set with the remaining records used as a test set. Decision tree classifiers were fit using the CART algorithm. Search of hyperparameters with k-fold cross-validation on the training set was used to guide model selection. A random forest classifier was trained on the same data for comparison. Area under ROC curve was computed against the test set.

Results: There were 57 chordomas and 16 chondrosarcomas with complete imaging for review. Of nine yes/no features labeled, nine were discriminatory in the single-tree model trained with all features. Important features: midline location—0.719; MRI T1 + C bubbly enhancement—0.169; chondroid matrix/calcifications on CT—0.074; MRI T1 hypo vs. hyperintense—0.037. Area under ROC curve was 0.86. A random forest classifier composed of 500 trees used 7 features with area under the ROC curve of 0.97 when applied to the test set.

Conclusion: This decision-tree model for radiographic differentiation of chordomas and chondrosarcomas may be useful for inexperienced physicians to establish a diagnosis and guide clinical decisions. A random forest classifier may overcome weakness due to over-fitting; however, interpretability is limited.