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

Abstract

In large clinical centers a small subset of patients present with hydrocephalus that requires surgical treatment. We aimed to develop a screening tool to detect such cases from the head MRI with performance comparable to neuroradiologists. We leveraged 496 clinical MRI exams collected retrospectively at a single clinical site from patients referred for any reason. This diagnostic dataset was enriched to have 259 hydrocephalus cases. A 3D convolutional neural network was trained on 16 manually segmented exams (ten hydrocephalus) and subsequently used to automatically segment the remaining 480 exams and extract volumetric anatomical features. A linear classifier of these features was trained on 240 exams to detect cases of hydrocephalus that required treatment with surgical intervention. Performance was compared to four neuroradiologists on the remaining 240 exams. Performance was also evaluated on a separate screening dataset of 451 exams collected from a routine clinical population to predict the consensus reading from four neuroradiologists using images alone. The pipeline was also tested on an external dataset of 31 exams from a 2nd clinical site. The most discriminant features were the Magnetic Resonance Hydrocephalic Index (MRHI), ventricle volume, and the ratio between ventricle and brain volume. At matching sensitivity, the specificity of the machine and the neuroradiologists did not show significant differences for detection of hydrocephalus on either dataset (proportions test, p > 0.05). ROC performance compared favorably with the state-of-the-art (AUC 0.90-0.96), and replicated in the external validation. Hydrocephalus cases requiring treatment can be detected automatically from MRI in a heterogeneous patient population based on quantitative characterization of brain anatomy with performance comparable to that of neuroradiologists.

Keywords: Brain MRIs; Convolutional neural networks; Deep learning; Hydrocephalus.

Fig. 1

Number of exams used in training and testing. Note that for the Screening Dataset there is overlap between the training set and testing set, as the training was performed using leave-one-out cross validation; i.e., for each test exam a different classifier was trained on the training data leaving out the one test exam

Fig. 2

Flowchart of the automated pipeline for machine detection of hydrocephalus

Fig. 3

Segmentation for two patients and representative volumetric features for all patients. A–C Segmentation for a non-hydrocephalus patient (NH) and hydrocephalus patients (H) from the Diagnosis Dataset, showing a sagittal, axial, and coronal views for the same two patients. D Distribution of three representative features extracted from these segmentations. Each point represents a patient (red: hydrocephalus, blue: non-hydrocephalus). Features are ratio of ventricle over extraventricular CSF volume (R_VC), volume of the temporal horns (V_H), and ratio of ventricle area over area of bounding box averaged over multiple coronal slices (E_2c; boxes are white rectangle in panels C1 and C2). Correlation coefficients between each pair of features are noted (*p < 0.05). Histograms of each feature are shown on the diagonal, with red and blue indicating hydrocephalus and non-hydrocephalus, respectively. Separability of each feature measured in Cohen’s d is also noted on the diagonal (**p < 0.001, Wilcoxon rank sum test, N = 240)

Fig. 4

Representative T1-weighted post-contrast images provided to neuroradiologists to make an imaging diagnosis of hydrocephalus requiring treatment. This is an example for a single exam

Fig. 5

Test-set performances of the machine and neuroradiologists (R1–R4) in detecting hydrocephalus. A and D Prediction of the clinical diagnosis of hydrocephalus requiring treatment in 240 exams (120 positive) in the Diagnosis Dataset. B and E Prediction of the majority readings in the 451 exams in the Screening Dataset; for each radiologist (R1–R3), a slightly different majority diagnosis serves as “ground truth,” hence different curves. C and F Prediction of the clinical diagnosis of hydrocephalus requiring treatment in 31 exams (15 positive) in the External Dataset

Fig. 6

Inter-rater agreement on 240 exams in the Diagnosis Dataset between the four neuroradiologists (R1–R4) and the surgical intervention truth labels (Clinical)