Automated Bone Age Analysis with Lossy Image Files

In our U.S. Department of Defense hospital system, pediatric endocrinology and radiology resources to evaluate bone age radiographs are limited. Our tertiary care center provides expert specialty support to remotely stationed beneficiaries at more than 30 Department of Defense medical facilities using a well-established, asynchronous, Health Insurance Portability and Accountability Act compliant system that allows for physician-to-physician teleconsultation. Up to 14% of these teleconsultations are for endocrinology assessment, many of which include bone age analysis. We sought to evaluate the feasibility of using an automated bone age analysis program using the file format most commonly provided to us, lossy JPEG image files saved at lower quality, to improve access to our consultation services.

All patients seen in the Tripler Army Medical Center pediatric endocrinology clinic, who were being evaluated for poor growth during the 2-month study period and who had a bone age film performed at Tripler Army Medical Center during that time, were eligible to have their deidentified bone age films analyzed. We imported lossy JPEG bone age image files from our hospital web viewer to BoneXpert, version 2.1, using a fully automated, custom built system that reconstructed each file's true resolution and then packaged the original image into a Digital Imaging and Communications in Medicine header. The original JPEG files were saved at 70% quality. Bone age readings were compared between our pediatric endocrinologists (ENDO), pediatric radiologists (RADS), and BoneXpert (BONE). Additionally, adult height prediction from ENDO and BONE were compared.

35 bone age images were evaluated over a 2-month period. Most patients were being evaluated for idiopathic short stature or growth hormone deficiency. Analysis of variance showed no significant differences in mean bone age readings between the 3 groups (mean bone age reading = 9.0, 9.1, and 9.1 years for ENDO, RADS, and BONE, respectively, p = 0.827). Mean (SD) differences between physician and software bone age readings were -0.09 (0.89) years (ENDO) and -0.03 (1.01) years (RADS). Mean difference for adult height predictions was only -0.2 cm (p = 0.806).

Automated analysis of lossy JPEG files of bone age images using the BoneXpert software appears to be feasible and accurate. Larger studies are needed to validate these results.