Bone age estimation using deep learning and hand X-ray images

Bones during growth period undergo substantial changes in shape and size. X-ray imaging has been routinely used for bone growth diagnosis purpose. Hand has been the part of choice for X-ray imaging due to its high bone parts count and relatively low radiation requirement. Traditionally, bone age estimation has been performed by referencing atlases of images of hand bone regions where aging-related metamorphoses are most conspicuous. Tanner and Whitehouse' and Greulich and Pyle's are some well known ones. The process entails manual comparison of subject's hand region images against a set of corresponding images in the atlases. It is desired to estimate bone age from hand images in an automated manner, which would facilitate more efficient estimation in terms of time and labor cost and enables quantitative and objective assessments. Deep learning method has proved to be a viable approach in a number of application domains. It is also gaining wider grounds in medical image analysis. A cascaded structure of layers can be trained to mimic the image-based cognitive and inference processes of human and other higher organisms. We employed a set of well known deep learning network architectures. In the current study, 3000 images were manually curated to mark feature points on hands. They were used as reference points in removing unnecessary image regions and to retain regions of interest (ROI) relevant to age estimation. Different ROI's were defined and used-that of rather small area mostly made up of carpal and metacarpal bones and that includes most of phalanges in addition. Irrelevant intensity variation across cropped images was minimized by applying histogram equalization. In consideration of the established gender difference in growth rates, separate gender models were built. Certain age range image data are far scarcer and exhibit rather large excursion in morphology from other age ranges-e.g. infancy and very early childhood. Many studies excluded them and addressed only elder subjects in later developmental stages. Considering infant age group's diagnosis demand is just as valid as elder groups', we included entire age ranges for our study. A number of different deep learning architectures were trained with varying region of interest definitions. Smallest mean absolute difference error was 8.890 months for a test set of 400 images. This study was preliminary, and in the future, we plan to investigate alternative approaches not taken in the present study.