Computer-aided Diagnosis For Radial Bone Age Assessment

Bone age is an important indicator of the physical development for adolescent children,widely used in areas of sports science,criminal justice and clinical medicine.Wrist can be more accurately describe the overall growth and development situation of bones.So international generally use the development of wrist as a standard evaluation for skeletal age.CHN05 scoring method is a standard method which used to assess bone age in China.However,the process of assessment needs professional personnel to learn proficient relative knowledge,besides the whole process is very cumbersome.Therefore,artificial bone age assessment has the disadvantages of strong subjectivity and low accuracy,so there is a growing demand for computer-aided system to automatically assess bone age.At present,there are two difficulties in computer-aided bone age assessment.First,the position,direction and size of wrist bone in CT image are unascertained,besides in later stage of development,two adjacent bones may appear the coincidence phenomenon,so there are difficulties in location and segmentation of wrist bone.Secondly,different growth and development characteristics described by human language in bone age standard are difficult to convert into computerized image features.In view of the above problems,this paper presents a method of computer-aided radius evaluation based on CHN05 scoring method.The automatic evaluation process of radius bone age is studied according to the characteristics of radius in CT images.The main contents of this paper are as follows:(1)A method of radial segmentation based on Constrained Local Model of multiple template is proposed.For changes of the shape in different levels of radial bone age are very intense,the Constrained Local Model is not easy to achieve convergence.So a multi-template Constrained Local Model algorithm is proposed,which is used to pre-classify radial images,then segment them with the model close to the true radial position.Experiments show that the algorithm has faster convergence speed and higher segmentation accuracy.(2)Considering the background of CT image contains only a flat color,this paper proposes a Haar Random Forest algorithm and uses it as a local detector for Constrained Local Model.According to experiments,the proposed algorithm has higher robustness and accuracy in segmenting the radial region in CT images than using PCA or SVM as a local detector.(3)According to the radius rating criteria of CHN05 scoring method,we select gray features and local shape features as representative features.At the same time,according to the characteristics of radial in CT images,global shape feature based on PCA and texture feature based on Scale Invariant Feature Transform have been extracted.Random Forest classifier is used to predict the radial bone age level,the parameters of classifier are optimized by experiments,and classification performance of the features extracted in this paper have been verified.