Bone Age Assessment based on SuperPoint Features

Convolutional neural network-based bone age assessment (BAA) algorithms have been researched in recent years. However, current state-of-the-art algorithms always have unacceptably large amount of parameters, so their performance will degrade when training on small scaled X-ray photographs datasets. In this paper, we aim to propose a light convolutional neural network to assess bone ages automatically. The proposed network consists of two components. The first module uses a pre-trained SuperPoint network to extract robust feature maps. The second module reduces the dimension of the feature maps and generate the final prediction by improving a classification network. Besides, a weighted loss function based on term-frequency index theory is designed to avoid over-fitting caused by uneven distribution of the training set. The parameters to be trained in our architecture is much fewer than state-of-the-art BAA networks. Experimental results on open-source X-ray dataset demonstrated that the proposed network can be trained on existing small scaled BAA datasets and achieves a comparable performance with those large scaled models for bone age assesment.