Automatic Bone Age Assessment of Radiographs using Deep Learning

One of the major factors that determine the growth of a child is bone age. Traditionally, this is determined by using techniques like the TW (Tanner - Whitehouse method) or the GP (Greulich Pyle method) on X-rays of the left hand. The X-rays are examined for any abnormality based on a standard set of regions of interest by a trained medical professional. Therefore, susceptibility to human error is extremely high which causes inconsistencies and often outputs noticeably inaccurate test results. In addition, this process is time-consuming which furthermore affirms the impracticality of using the traditional method. To combat this, deep learning architectures like Convolution Neural Networks (CNN) and their modified counterparts are used to produce significantly more accurate results, in less time. In this paper, we use one such architecture called Xception. This architecture fundamentally replaces the standard Convolution operation with a much more efficient operation called Depthwise Separable Convolution, which in turn drastically reduces the time taken to build a model. Apart from being computationally quick, an exceptional deep learning model must also give accurate results, this is made possible by training a model on an enormous dataset. In this paper, we use a dataset of left-hand Radiographs provided by RSNA. The application of an efficient activation function also contributes to making a better model, in this paper we use two activation functions namely ReLU and Swish to demonstrate the significance activation functions play on the accuracy of the model. The results obtained by our paper indicate that the swish activation function outperforms ReLU in deeper convolution, providing us with 0.183 years MAE compared to the 0.2414 years MAE given by ReLU.