Deep Learning-Based Bone Age Prediction Using VGG16: An Attention-Driven Approach for Pediatric Radiograph Analysis

The research proposes a deep learning model based on the VGG16 CNN for the estimation of bone age using radiographs of the hand. The model performs age regression in low-resolution radiographs (384×384) with several data augmentation techniques and an integrated attention mechanism that enhances the accuracy of the model. Furthermore, the dataset was divided into age groups to classify and evaluate the model. The model is trained based on these categories. The best performance of the proposed model was an overall classification accuracy of 95%. Overall, the model precision, recall, and F1-scores were above 94% across different age categories of interest. For example, the precision ranges from 0.94 to 0.97, and recall also ranges from 0.94 to 0.97 across the age ranges. An attention mechanism was employed, improving the ability of the model to narrow its focus on relevant regions of the radiographs, hence contributing a lot to the resultant high accuracy. The training process has shown quite steady improvement in both training and validation loss with a similar MAE. Considering the low resolution of images which the model dealt with, good performance of mean absolute error could be achieved at 14.48 months, hence being good enough for bone age prediction. Further work on higher-resolution images with advanced normalization techniques needs to be done to improve this model.