Adaptive Critical Region Extraction Net via relationship modeling for bone age assessment

Abstract

Bone age assessment is an efficient clinical practice to diagnose children’s endocrine, genetic, and growth disorders. Guided by clinical diagnostic methods, local critical regions cropped from original hand radiography images have significant biological identification meaning in skeletal maturity analysis. However, existing methods crop local regions without maintaining the relationship between cropped regions and the global image, which results in discarding essential information. To address this problem, we propose an Adaptive Critical Region Extraction Net (ACRE-Net) for bone age assessment, which can automatically locate local critical regions from global images and calibrate features by relationship modeling. Unlike existing two-stage methods, which crop global images in the first stage and analyze cropped parts in the second stage separately, our ACRE-Net maintains the relationships between global images and cropped regions and further integrates such relationship information into relationship-guided features. With the relationship-guided features, we propose a novel cross-graph feature calibration module which is consisted of a bigraph structure to supplementary and calibrate features extracted from cropped parts by also modeling the relationship between cropped parts. Meanwhile, we offer a novel localization strategy in the patch-based region localization module, which aggregates and leverages multi-scale attention maps to locate patches of interest as critical regions to be cropped. In practical application, our ACRE-Net can free clinicians from time-consuming manual annotations on local critical regions and automatically assess bone age through radiography images. Extensive experiments illustrate that our ACRE-Net achieves state-of-the-art performance with Mean Absolute Error (MAE) 3.78 months on the public RSNA2017 dataset.