Deep Learning Achieves Neuroradiologist-Level Performance in Detecting Hydrocephalus Requiring Treatment

Abstract

PurposeWe aim to develop automated detection of hydrocephalus requiring treatment in a heterogeneous patient population referred for MRI brain scans, and compare performance to that of neuroradiologists.Materials and MethodsWe leveraged 496 clinical MRI brain scans (259 hydrocephalus) collected retrospectively at a single clinical site from patients aged 2–90 years (mean 54) referred for any reason. Sixteen MRI scans (ten hydrocephalus) were segmented semi-automatically in 3D to delineate ventricles, extraventricular CSF, and brain tissues. A 3D CNN was trained on these segmentations and subsequently used to automatically segment the remaining 480 scans. To detect hydrocephalus, volumetric features such as volumes of ventricles and temporal horns were computed from the segmentation and were used to train a linear classifier. Machine performance was evaluated in a diagnosis dataset where hydrocephalus was confirmed as requiring surgical intervention, and compared to four neuroradiologists on a random subset of 240 scans. The pipeline was tested on a separate screening dataset of 205 scans collected from a routine clinical population aged 1–95 years (mean 56) to predict the majority reading from four neuroradiologists using images alone.ResultsWhen compared to the neuroradiologists at a matched sensitivity, the machine did not show a significant difference in specificity (proportions test, p > 0.05). The machine demonstrated comparable performance in independent diagnosis and screening datasets. Overall ROC performance compared favorably with the state-of-the-art (AUC 0.82–0.93).ConclusionHydrocephalus can be detected automatically from MRI in a heterogeneous patient population with performance equivalent to that of neuroradiologists.Summary statementA two-stage automated pipeline was developed to segment head MRI and extract volumetric features to accurately and efficiently detect hydrocephalus that required shunting and achieved performance comparable to that of trained neuroradiologists.Key PointsWe developed a state-of-the-art 3D deep convolutional network to perform fully automated segmentation of the ventricles, extraventricular cerebrospinal fluid, and brain tissues in anisotropic MRI brain scans in a heterogeneous patient population.Volumetric features extracted from anatomical segmentations can be used to classify hydrocephalus (which may require neurosurgical intervention) vs. non-hydrocephalus.When tested in an independent dataset, the network achieved performance comparable to that of expert neuroradiologists.