Automated vision system for skeletal age assessment using knowledge based techniques

This paper presents a knowledge-based automated vision system to segment bones in a child's hand radiograph image, and to determine growth progress using decision theoretic approaches. A hierarchical knowledge-based localisation scheme is used to localise bones in the hand radiograph image. Bone contour detection is then implemented with further knowledge represented by Active Shape Models (ASM). Hence a set of parameters is generated to describe the bone contour shape. The bone image is parameterised to describe its texture which is correlated to growth age. Regression and Bayesian methods are then used to model the characteristics of the most correlated shape parameters to the growth age as well as texture parameters in a training set. The models are finally applied to test images to estimate their bone ages. The Bayesian methods result in an 8.93% average relative error.