Detection of idiopathic normal pressure hydrocephalus on head CT using a deep convolutional neural network

Idiopathic normal pressure hydrocephalus (iNPH) is an underrecognized cause of dementia, with reasons for underdiagnosis including symptomatic overlap with other neurologic disorders and difficulty in distinguishing the disproportionate ventriculomegaly of iNPH from generalized volume loss on cross-sectional imaging. In response to this problem of underdiagnosis, we developed a convolutional neural network (CNN) to detect iNPH on head CT. In this retrospective study of 358 patients with head CTs acquired between 1997 and 2020, a CNN was trained to identify iNPH. The iNPH cohort utilized a clinically derived ground truth based upon EMR metadata meeting Japanese Society of Normal Pressure Hydrocephalus 2nd Edition criteria for definite iNPH. The non-iNPH cohort included matched control patients. Statistical analysis included evaluation of AUROC for the test partitions. The cohort included 80 iNPH and 278 non-iNPH patients identified for inclusion. Test partition performance demonstrated 100% sensitivity [95% CI: 100%, 100%] and 89% specificity [95% CI:78%,97%], with four false positives, zero false negatives, and a 0.96 AUROC [95% CI:0.89,0.99]. We therefore demonstrated that a CNN utilizing a clinically derived ground truth can identify iNPH on head CT. Our model has the potential to be used in clinical practice as a screening tool to assist in the detection of iNPH in settings in which a high volume of head CTs is performed, and for a variety of indications, such as within the emergency department, thus potentially identifying patients who may benefit from referral to neurology or neurosurgery for further evaluation.