Purpose or Case Report: The shortage of pediatric datasets challenges the development and evaluation of safe and effective artificial intelligence and machine learning (AI/ML)-based medical devices for pediatrics. This is the case for computer-aided triage (CADt) of intracranial hemorrhage (ICH) where the incidence of ICH in the pediatric population is lower than adults thus limiting pediatric data for evaluations. Here we introduce tools for generating synthetic adult and pediatric non-contrast computed tomography (NCCT) datasets with ICH modeled from limited real data to address the shortage of pediatric data for ICH device assessments.
Methods & Materials: Epidural, subdural, and intraparenchymal hemorrhages of varying volume and location were modeled using anatomic constraints and inserted into MR-based pediatric brain atlases with a simulated local mass effect. These ground truth head models were imaged with a virtual CT scanner with a head scanning protocol and varying dose. Synthetic dataset quality was assessed with a top performing model from the 2019 RSNA ICH Detection Challenge. The detectability of synthetic ICH was compared to that of an independent real patient dataset of 75 NCCT exams (ICH incidence 50%) including 28 pediatric cases. Patient-level performance for the real and synthetic datasets was evaluated with the area under the curve (AUC) of the receiver-operator characteristic (ROC) curve.
Results: Three synthetic datasets with an average of 96 cases each (ICH incidence 75%) were created using pediatric brain atlases ranging from 4.5-18.5 years old. The synthetic data AUC was 0.929±0.011 compared to the real patient test set AUC of 0.94, supporting the use of synthetic data to complement limited real patient data.
Conclusions: We demonstrate the use of synthetic data for evaluating standalone CADt detection of ICH in adults and pediatrics while varying patient, lesion, and acquisition parameters. We show a synthetic dataset achieved a similar AUC as a real dataset, suggesting that synthetic data could be used to complement real datasets by alleviating age imbalance without the added cost of acquiring more real data. Further research is needed in assessing the quality of synthetic data to better understand the role it can play in medical device evaluations.