Automatic assessment of bone age using statistical models of appearance and Random Forest Regression voting

This work addresses the problem of automatic estimation of bone age in children and young adults. Bone age assessment is important for diagnosing and monitoring growth and endocrine disorders. We have constructed a system which uses Statistical Models of Shape and Appearance to locate bones in a radiograph and to predict bone age. By analyzing the performance on a data-set of about 600 digitized radiograph of normal children we show how Random Forest Regression voting in a Contrained Local models framework (RFRV — CLM) are sufficient to initialize an automatic registration algorithm. We built global models of whole hand and local models of individual bones. We used the same RFRV — CLM models to locate salient bones of the hand. We improved our age estimation results by using multiple local age group models and multiple local age predictors. We obtained an accuracy of mean point-to-point errors of 0.87mm on sparse points placement for initialization of automatic registrations. Our Bone age assessment methodology achieved an accuracy of mean absolute error of 0.41±0.02 years and 0.47 ± 0.03 years for female and male respectively.