Research On Several Key Technologies Of Bone Age Assessment Based On Deep Learning

Bone age is the biological age obtained by comparing the bone development level of children with the standard bone development.It plays an important role in the analysis and diagnosis of growth disorders and hereditary diseases,the determination of the age of criminal responsibility and the selection of competitive sports talents.The manual assessment method is based on the subjective judgment of doctors on the maturity of several specific bones in the children’s left hand X-ray image.This process is time-consuming,labor-intensive,and prone to large errors.Therefore,automatic bone age assessment has gradually received attention from researchers.In the new era of AI infrastructure,automatic bone age assessment based on deep learning has developed rapidly.Its stable,fast and efficient features can effectively reduce medical costs and improve the accuracy of diagnosis,which has broad application prospects.In order to promote further clinical application,the following three problems need to be solved: first,how to effectively use different characteristic information of hand bone to improve the assessment performance,while ensuring that the network is not too complex;The second is how to carry out effective training when the hand bone data set is incomplete and there are human labeling errors;The third is how to effectively carry out automatic bone age assessment under different populations,different equipment,different scanning conditions and without human data annotation.Aiming at the appeal problem,this paper studies several key technologies of automatic bone age assessment based on deep learning from the perspective of fine-grained image classification,label distribution learning and unsupervised domain adaptation.The main work is as follows:(1)A fine-grained progressive multi-branch bone age assessment method.Constructing a multi-branch network by sharing weights to divide a picture into inputs containing different granularity information,and adopt a progressive training strategy to gradually integrate the global and local features of the hand bone,so as to improve the bone age assessment performance without increasing the complexity of the network.(2)A bone age assessment method based on label distribution learning.By using the normal distribution to establish fuzzy relationships between labels,the bone age labels are re-expressed to expand the feature expression of missing categories by utilizing the correlation and orderliness of adjacent bone age categories,thus expanding the dataset.By combining prediction outputs from different training periods to guide the training process,the potential probability of correct categories can be effectively improved,reducing the impact of human annotation errors and enhancing the accuracy and robustness of bone age assessment.(3)A bone age assessment method based on unsupervised domain adaptation.Based on the concept of domain distribution difference,the concepts of "intra-class" difference and "inter-class" difference are proposed.Through class-aware sampling,discriminant feature learning and domain alignment are jointly performed,and bone age assessment is performed on target domain samples without labels,so that the network has stronger generalization and robustness.