Advances in Deep Learning-Based Bone Age Assessment Research

Abstract

In clinical and forensic medicine, bone age assessment (BAA) is crucial for tracking children's growth and development as well as diagnosing and identifying diseases. While statistical machine learning and mapping and expert discrimination are two examples of traditional BAA techniques, deep learning (DL)-based BAA techniques are progressively becoming more popular as deep learning technology advances. The use of convolutional neural networks in BAA is explained in length in this study, along with information on their development history, particular applications in the segmentation and regression stages of BAA, and related models. Currently, deep learning bone age assessment methods mainly include end-to-end assessment method and region of interest-based assessment method, and in regression, it has also experienced a stage from early exploration to gradual development to rapid development. However, the existing methods have problems such as insufficient attention to image detail information, partial reliance on manual annotation, inaccurate assessment results, and lack of in-depth study of local bone features. In the future, efforts should be made to make better use of image detail information, explore weakly supervised learning methods, conduct in-depth research on local skeletal features, and combine multimodal data to improve the performance of BAA and promote the wide application of this technology in clinical and other fields.