326 Prediction of Shunt Malfunction Using Automated Ventricular Volume Analysis and Radiomics

INTRODUCTION: The assessment of ventricle size is crucial in diagnosing hydrocephalus and detecting shunt malfunctions. Current methods primarily involve 2 dimensional measurements, which can be time-consuming and subject to inter-observer variability. METHODS: We identified patients that underwent placement or revision of a VP shunt between January of 2018 and August of 2020 and collected head CTs from patients which were imaged with a functional shunt, a shunt malfunction, or prior to the placement of their shunt. A convolutional neural network (CNN) was used to automate ventricle segmentation. The automated segmentations were verified and hand corrected when necessary. For each patient, the ventricular volume was compared against a reference age and sex normative dataset and radiomics analyses were performed on the pathological and a normal dataset. The imaging data was combined with clinical features, and used to train several classifiers to identify patients with a malfunctioning shunt or hydrocephalus. RESULTS: 145 head CTs from 66 patients were segmented. Automated segmentation required correction in 24% of the scans due to inadequate segmentation. Volumetric analysis alone yielded an AUC of 0.772. The addition of radiomic data achieved an AUC of 0.929 on the testing dataset with XGBoost. When radiomic features were combined with age and gender, XGBoost achieved an AUC of 0.946 on the testing dataset. An SVM was able to achieve an AUC of 0.925 on testing dataset with normal and pathologic scans. CONCLUSIONS: Automated ventricle segmentation using CNNs combined with radiomic analysis can assist the detection of hydrocephalus and shunt malfunctions. It offers a time-efficient alternative to manual segmentation, reduces inter-observer variability, and holds promise in improving patient outcomes by facilitating early and accurate diagnosis of hydrocephalus/shunt malfunction.