Dense Cross-Scale Deep Learning with Ordinal Skeletal Maturity Modelling for Pediatric Bone Age Estimation

Abstract

Bone age analysis is a vital process in pediatrics that entails evaluating the degree of development and detecting abnormalities in a child's growth and development using images of the hand. The existing deep learning models for bone age estimation have considered the problem as a continuous regression problem and utilized basic multi-scale feature integration, which results in the loss of structural consistency and ordinal progression of bone development. This creates a major gap in feature learning in a cross-scale perspective and modelling of ordinal skeletal maturation, making age estimation less reliable and inadequate. Given that bone maturity is sequential, an approach based on ordered skeletal maturation can be more appropriate as opposed to regular regression-based age estimation since it takes into account the structural connections between consecutive growth levels. This study introduces a new DCSF-DCNN method combined with ordinal bone maturation learning and ParaU-Net segmentation. Preprocessing involves image resizing and min-max normalization, whereas the proposed ParaU-Net accurately segments carpal bones, metacarpals, and phalanges to eliminate irrelevant information. The DCSF module helps in improving hierarchical feature consistency through dense cross-scale refinement, while the ordinal learning module models bone maturation order for improved age prediction. An experiment conducted using a publicly available dataset of 14,236 hand radiographs in the RSNA data showed that the proposed DCSF-DCNN method has outperformed the baseline models, achieving a MAE of 1.27, RMSE of 1.58, MSE of 2.40, and MAPE of 1.92. This means that the method is better at age prediction than baseline algorithms like Residual CNN, CNN, and DCNN as it achieves an accuracy of 96%.