Automated Bone Age Assessment in Pediatric Radiology Using MobileNet: A High-Accuracy Deep Learning Approach for Balanced Age Group Predictions

Grading of bone age is thus highly relevant for the diagnosis of growth disorders in children. Conventionally, this has been done by manually analyzing the radiograph using an atlas either by Greulich-Pyle or Tanner-Whitehouse, a method highly time-consuming and prone to inter-observer variability. Based on this fact, a light convolutional neural network was used-Mobilenet-for automated prediction of bone age from radiographs of the hands of pediatric patients. Because of its depthwise separable convolutions, MobileNet is suitable for deployment on resource-constrained devices. This finally opens up the possibility of deploying models in real-time clinical use. Our model was therefore trained using a dataset labeled per age, ranging from 0 to 168 months, achieving an overall accuracy of 93%. The following classification report showed that the performance was consistently high within all the age groups, averaging 0.93 for precision, recall, and F1-scores. In addition, age classes like 0–24 months and 49–72 months both had a precision value close to 0.95 and recall also close to 0.95, indicating that the model is robust for early and mid-childhood categories with seldom growth rates. MobileNet architecture developed for a wide range of applications means that this model responds to a current need for a reliable, accessible tool regarding pediatric bone age assessment. The performance of MobileNet in our study is comparable to that of deeper networks while offering advantages computationally and in feasibility of deployment. This study identifies that MobileNet has the potential to help clinicians with fast and accurate assessment of bone age to eventually enable early detection and treatment of pediatric growth abnormalities. Testing with a more diverse set of data could probably improve generalizability across different populations.