Prediction of Hydrocephalus in Colloid Cysts Using Artificial Intelligence

Abstract

Objective: We aim to train neural networks to predict hydrocephaly in patients with colloid cysts based on T2-weighted magnetic resonance imaging radiomics. Methods: This study included 40 cases with a colloid cyst; the mean age was 54.08 ± 16.57 years, and 25 (62.5%) were women. Two observers segmented cysts on axial T2-weighted MRI and evaluated conventional features. Predictors were radiomics (n = 851) and conventional features (n = 12). Feature selection was based on coefficient variance (CoV), variance inflation factor (VIF), and least absolute shrinkage, and a selection operator regression analysis. The outcome was identified as hydrocephaly. Models were developed with artificial neural networks (ANN) for 3 different diag- nostic prediction models. The first model included radiomics features; the second model included conventional features; and the third model included all of the features. Artificial neural network performance was presented as an area under curve (AUC) and the receiver operating characteristic curve (ROC) and accepted as successful if the AUC > 0.85 and p-value < .01. Results: By using CoV and VIF analysis, 49 features were found to be stable. Radiomics predict hydrocephaly with AUC = 0.88, sensitivity: 92%, specificity: 97%. Conventional features predict hydrocephaly with AUC = 0.87, sensitivity: 82%, and specificity: 93%. Third model (radiomics + conventional) AUC was 0.99, sensitivity: 91%, and specificity: 100% (all p-values < .001). Conclusion: This study was successful in training neural networks that can predict hydrocephaly in patients with colloid cysts.