Using the 2017 Pediatric Bone Age Challenge organized by the Radiological Society of North America, we attempted to accurately determine skeletal age in a curated data set of pediatric hand radiographs. The primary evaluation measure was the mean absolute distance (MAD) in months, which was calculated as the mean of the absolute values of the difference between the model estimates and those of the reference standard, bone age. Using a data set consisting of 14,236 hand radiographs (12,611 training set, 1,425 validation set, 200 test set), we implemented a convolutional neural network architecture based on VGG-16 in order to estimate the bone age. Our results showed satisfactory results and proved that a computational model can perform fast and accurate calculations and can be a possible replacement for current gold-standard Greulich and Pyle method, which consists of an atlas in which bone age is assessed by comparison of a patient's left hand x-ray with one of the closest standard x-rays in a reference database.
We used an architecture inspired by VGG-16 which consisted of six VGG blocks, each with two convolutional layers and corresponding ELU, batch normalization, and max pooling layers. We followed this with three dense layers with dropout to obtain our model’s prediction. We trained over 3 epochs using the Adam optimizer with a learning rate of 2e-5, with a batch size of 3 and an image size of (800,600).
Currently, our model reaches a testing accuracy of 33.1341 months (measured through mean absolute error)