Bone Age Assessment Based on Salient Object Segmentation

Predicting bone age has become pivotal to measure developmental disorder and endocrinology in children. Traditional procedures such as Greulich–Pyle and Tanner-Whitehouse have been used intensively to measure bone age. Recently, several proposed deep learning algorithms show high statistical predictions compared to the traditional assessment practice. However, the commonly used hand x-ray data in bone age assessment, RSNA, suffers from multi-level intensities and unneeded marks in backgrounds. This leads to considering inconsistent features in training models and reduced efficiency. Attention map is introduced here to visualize model attentions on unwanted features. To tackle this flaw and to elevate the assessment quality, we propose a framework technique to segment hands from the noisy background before training the model. We utilize a pretrained salient object detection network, U² -net, to detect crucial parameters in RSNA samples. Presented results show notable improvements in segmenting hands from x-ray samples and a significant statistical prediction in bone age assessment based on the InceptionV3 algorithm.