Bone age estimation from children's hand-wrist radiology images using deep learning

Bone age estimation is used in diagnosing endocrine and metabolic disorders in child development, in forensic medicine cases, in registering the accuracy of the age declared in the identities of individuals. In this thesis, a deep learning-based application was carried out for the determination of bone age from the hand-wrist radiology images of girls and boys. Bone age determination is made by examining the hand-wrist radiology images of specialist physicians by eye procedure. A well-trained deep learning model can be used to effectively predict/classify small details that the human eye has difficulty seeing or interpreting. In this study, the publicly available RSNA dataset was used. For the study, a total of 2321 X-ray hand-wrist images, 1206 girls and 1115 boys, between the ages of 1-7 were examined. Images were analyzed in three different deep learning architectures, ResNet, DenseNet and EfficientNet. Feature vectors of each image were obtained with three deep learning models and these vectors were classified with support vector machines. Images were analyzed in four different ways. Raw images were used in the first review. In the second, unwanted parts of the image were removed. In the third method, data augmentation was applied due to the data imbalance between age groups. In the fourth method, edge extraction method was applied to the preprocessed images. The prediction accuracy of preprocessed images was approximately 8% better in girls and 16% better in boys than raw images. In the same way, when a balanced data number is reached with data augmentation, an accuracy increase of approximately 21% for girls and about 20% for boys compared to pre-processed images was obtained. In the estimation made after the edge extraction algorithm was applied, the success was lower than the data augmentation method. In all three deep learning models, values close to each other have been reached in different age and gender groups. It was seen that all three models were able to extract features effectively from the images and as a result, they achieved a successful prediction accuracy. As a result, it has been seen that preprocessing of hand-wrist X-ray images, large data set and creating a balanced data set make a great contribution to the production of successful estimation results of convolutional neural networks.