Bone Anomaly Detection by Extracting Regions of Interest and Convolutional Neural Networks

The most suitable method for assessing bone age is to check the degree of maturation of the ossification centers in the radiograph images of the left wrist. So, a lot of effort has been made to help radiologists and provide reliable automated methods using these images. This study designs and tests Alexnet and GoogLeNet methods and a new architecture to assess bone age. All these methods are implemented fully automatically on the DHA dataset including 1400 wrist images of healthy children aged 0 to 18 years from Asian, Hispanic, Black, and Caucasian races. For this purpose, the images are first segmented, and 4 different regions of the images are then separated. Bone age in each region is assessed by a separate network whose architecture is new and obtained by trial and error. The final assessment of bone age is performed by an ensemble based on the Average algorithm between 4 CNN models. In the section on results and model evaluation, various tests are performed, including pre-trained network tests. The better performance of the designed system compared to other methods is confirmed by the results of all tests. The proposed method achieves an accuracy of 83.4% and an average error rate of 0.1%.