BonCC:A Lightweight TW3-based Bone Age Assessment Coordinate Classification Model

Bone age is an important indicator of children's growth and development, and using deep learning to rapidly and accurately assist in the assessment of children's bone age has become a current popular trend. There are many methods for the automatic assessment of bone age that draw from the Tanner-Whitehouse 3 (TW3) method in bone age assessment, which is a two-stage method of "keypoint detection—skeletal staging prediction." However, existing research often focuses on the design of the staging model, neglecting the importance of keypoint detection. In fact, the accuracy of keypoint detection directly influences subsequent staging outcomes, and thus this is a weak point in current studies. This paper proposes structure called BonCC based on Simple Coordinate Classification (SimCC), which integrates Feature Pyramid Network (FPN) and Guided Attention Unit (GAU) for hand skeletal keypoint localization. The overall method uses mainstream Backbone such as ResNet-50 for feature extraction and inputs the feature map into the feature pyramid to fuse semantics of different levels. It then elegantly incorporates self-attention mechanisms into the network structure through the GAU, and finally, the SimCC method is used to transform the 2D coordinate regression into a classification problem for two coordinate axes, avoiding the subsequent computational processing of traditional pose detection methods. Experiments conducted on the Radiological Society of North America (RSNA) bone age dataset demonstrate that BonCC excels in the precision of hand keypoint detection while still achieving a lightweight design, especially notable in low-resolution situations.