Purpose or Case Report: To assess whether a convolutional neural network (CNN) can be trained via transfer learning to accurately diagnose germinal matrix hemorrhage-intraventricular hemorrhage (GMH-IVH) on cranial ultrasound (CUS) by comparing its performance to that of a board-certified radiologist. Secondary endpoints will include the ability of CNN to grade GMH-IVH severity, identify GMH-IVH on MRI, and assess low-grade GMH-IVH on CUS that cannot reliably be seen by a radiologist.
Methods & Materials: In this IRB approved retrospective study, we reviewed records from our children hospital’s radiology database. Consecutive patients, 6 months in age or younger, from January 2009 to September 2019, were screened for cranial ultrasounds performed at our institution with or without subsequent brain MRI. Patients with inadequate or incomplete ultrasound examinations due to a poor acoustic window from a closing anterior fontanelle were excluded.

Types of data collected were as follows: patient characteristics, such as age, gender, diagnosis, and indication for imaging referral; and imaging data, including right and left GMH-IVH grade on ultrasound based on the 1978 classification proposed by Burstein, Papile et al. and presence of GMH-IVH on MRI.

A total of 200 CUS were retrospectively utilized for the training and testing of a CNN binary classifier. Key sagittal images at the level of the caudothalamic groove were obtained from 100 patients with GMH-IVH and from 100 patients without hemorrhage. 20 cases were randomly allocated from the total for validation and an additional 20 for testing. We utilized transfer learning (VGG16) and data augmentation in our algorithm due to the low sample size. The model was then assessed using the test dataset with area under the Receiver Operating Characteristic curve and accuracy.
Results: Preliminary analysis using this small cohort suggests an accuracy in the 70-80% range with similar values for the area under the ROC curve.
Conclusions: A CNN trained on a small set of images with data augmentation is able to detect GMH-IVH on CUS with relatively good accuracy at superhuman speeds. We predict that further analysis with segmentation using a larger cohort will improve accuracy. This proof of concept may apply to other CUS diagnoses with the potential to create a neural network for grading GMH-IVH.