Web-based bone age assessment by content-based image retrieval for case-based reasoning

Abstract

Maturity estimation by radiological bone age assessment (BAA) is a frequent task for pediatric radiologists. Following Greulich and Pyle, all hand bones are compared with a standard atlas, or a subset of bones is examined according to Tanner and Whitehouse. We support BAA comparing the epiphyses of a current case to similar cases with validated bone age by content-based image retrieval (CBIR). A web-based prototype case-based retrieval system for BAA was developed and is publicly available. Hand radiographs from the USC database or user uploads may be retrieved by image-based query. The ten best matching cases for each epiphysis are retrieved by CBIR and displayed with their BAA, similarity score, and the derived age estimate. The similarity is approximated by cross-correlation. The USC hand database includes 1,101 cases comprising four ethnic groups of both genders between zero and 18 years of chronological age with radiographs and two annotated BAA. The USC image data have been enriched by marking the epiphyseal centers between metacarpals and distal phalanges. Leave-one-out experiments yielded a mean error rate of 0.99 years and a standard deviation of 0.76 years in comparison with the mean USC-BAA. The research prototype enables radiologists to judge their agreement based on similarity of retrieved cases and the derived age. CBIR provides support to the radiologist with a second opinion for BAA. Self-explanatory web applications can be established to support workflow integration. Enhancements in similarity computation and interface usability may further improve the system.