Background: Clinicians and machine classifiers reliably diagnose pilocytic astrocytoma (PA) on magnetic resonance imaging (MRI) but less accurately distinguish medulloblastoma (MB) from ependymoma (EP). One strategy is to first rule out the most identifiable diagnosis.
Objective: To hypothesize a sequential machine-learning classifier could improve diagnostic performance by mimicking a clinician's strategy of excluding PA before distinguishing MB from EP.
Methods: We extracted 1800 total Image Biomarker Standardization Initiative (IBSI)-based features from T2- and gadolinium-enhanced T1-weighted images in a multinational cohort of 274 MB, 156 PA, and 97 EP. We designed a 2-step sequential classifier - first ruling out PA, and next distinguishing MB from EP. For each step, we selected the best performing model from 6-candidate classifier using a reduced feature set, and measured performance on a holdout test set with the microaveraged F1 score.
Results: Optimal diagnostic performance was achieved using 2 decision steps, each with its own distinct imaging features and classifier method. A 3-way logistic regression classifier first distinguished PA from non-PA, with T2 uniformity and T1 contrast as the most relevant IBSI features (F1 score 0.8809). A 2-way neural net classifier next distinguished MB from EP, with T2 sphericity and T1 flatness as most relevant (F1 score 0.9189). The combined, sequential classifier was with F1 score 0.9179.
Conclusion: An MRI-based sequential machine-learning classifiers offer high-performance prediction of pediatric posterior fossa tumors across a large, multinational cohort. Optimization of this model with demographic, clinical, imaging, and molecular predictors could provide significant advantages for family counseling and surgical planning.
Keywords: Artificial intelligence; Ependymoma; Machine learning; Medulloblastoma; Pilocytic astrocytoma; Posterior fossa tumors; Radiomics.
© Congress of Neurological Surgeons 2021.