Purpose or Case Report: Knowing the gestational status of a neonate (premature or full-term) impacts a pediatric radiologist’s ability to generate a reasonable differential diagnosis of neonatal diseases. Though crucial, this information is often unavailable at the time of the babygram interpretation. Conventional methods based on measuring the clavicular-pubic length (CPL) and the transverse diameter of the chest (TDC) from a babygram have been proposed as reliable estimators of a neonate’s gestational age. In this study, we aimed to compare these two conventional methodologies to that of a deep-learning (DL) model in determining a neonate’s gestational status.
Methods & Materials: We conducted a computerized search of imaging reports and medical records from a large tertiary children's hospital to identify babygrams of neonates ≤48 hours-old with known gestational ages (5/2012--4/2022). Only one babygram per neonate was included in the database. Exclusion criteria were suboptimal positioning, limited imaging field-of-view, and imaging report of skeletal dysplasia. We measured the CPL and TDC of the babygrams and converted them to gestational ages based on the formulas reported in the literature. Estimated gestational ages ≥37 weeks were considered as full-term; otherwise as premature. Next, we employed a DL model (ResNet-34) that minimized the cross-entropy loss function to binary classify a babygram as premature or full-term. We used a 5-fold cross-validation procedure to determine the accuracy statistics of this model. These three methodologies were compared to one another via diagnostic performance statistics.
Results: The curated database consisted of 615 babygrams. The average gestational age of these neonates was 37 weeks (SD=4 weeks; range=23-42 weeks). The DL model estimated the gestational status of the neonates with 85% sensitivity, 63% specificity, 80% positive predictive value (PPV), 73% negative predictive value (NPV), and 77% accuracy. The sensitivity for the methods based on CPL and TDC was 70% and 73%, specificity was 83% and 77%, PPV was 87% and 84%, NPV was 62% and 63%, and accuracy was 75% and 74%, respectively.
Conclusions: The accuracy performances of the three methodologies were comparable. Despite the small size database, the DL model showed encouraging results. With further increases in the size and diversity of the database, the DL model will only improve its performance and might serve as a tool for radiologists in automatically determining the gestational status of a neonate based on a babygram.