Fully automated bone age assessment using bag of features on hand radiograph images

Bone age assessment (BAA) considered an essential task is performed on a daily basis in hospitals all over the world with the main indication being skeletal development in growth-related abnormalities. The manual methods for BAA are time consuming and subjective, which leads to imprecise and less accurate results. Thus, rendering the automated BAA more favorable. The purpose for BAA is to compare the measurement to chronological age so as to: Monitoring treatments and predict final adult height, observe the development for the skeleton and diagnose growth disorders, and to confirm age claims for children made by asylum seekers. Automated bone age assessment (ABAA) systems have been developed, none of these systems have been accepted for clinical use because there is a lack of agreement concerning the accuracy of bone age methods which is acceptable for a clinical environment. Most of the previously proposed methods for bone age assessment were tested on private x-ray datasets or do not provide source code, thus their results are not reproducible or usable as baselines. The previously proposed methods suffer from two main limitations: first, most of the methods operate only with x-ray scans of Caucasian subjects younger than 10 years, when bones are not yet fused, thus easier than in older ages where bones (especially, the carpal ones) overlap. Second, all of them assess bone age by extracting features from the bones either epiphyseal-metaphyseal region of interest (EMROIs) or carpal region of interest (CROIs) or both of them commonly adopted by the Tanner and Whitehouse (TW) or Greulich and Pyle (GP) clinical methods, thus constraining low-level (i.e., machine learning and computer vision) methods to use high-level (i.e., coming directly from human knowledge) visual descriptors. The analysis of bone age assessment becomes more complex when the bones are nearing maturity, when most of the bone would have merged, while some might overlap. The existing model-based approaches in the literature often reduce the region of interest (ROI) drastically to simplify the image analysis process, but this often leads to inaccurate and unstable results. Any system that attempts to automate skeletal assessment in an accurate manner will need to consider the entire span of the hand radiograph. Reduced ROI leads to inaccurate and unstable results. This semantic gap usually limits the generalization capabilities of the devised solutions, in particular when the visual descriptors are complex to extract as in the case of mature bones. A novel machinelearning framework presented, aimed at overcoming these problems by learning visual features. The proposed framework is based on speeded-up robust features (SURF) combined with bag of features (BoF) models to quantize features computed by SURF. Support vector machines (SVM) are used to classify the simplified feature vectors, extracted from hand bone x-ray images. Overall 745 images were obtained, 472 images for males, 273 images for females, most of them belong to chronological ages centered around 15 to 18 years. The proposed framework allows achieving classification results with an average accuracy of 99%, mean absolute error 0.012 for the 17 years and 18 years for the male gender with the SURF and BoF approach. In the female model, the age range from 0 to 7 years are excluded, and in the male model from 0 to 8, because of the limited amount of data that obtained, the female model range starts from 8 years to 18 years with classification average accuracy of 82.6%. The male model range starts from 9 years to 18 years with classification average accuracy of 85%.

Text FK 2019 4 UPMIR.pdf Download (1MB)

Actions (login required)

View Item