Assessing the Bone Age of Children in an Automatic Manner Newborn to 18 Years Range

Area of Science:

• Radiology
• Computer Vision
• Pediatric Endocrinology

Background:

• Bone age assessment (BAA) is crucial for diagnosing pediatric growth disorders but is time-consuming for radiologists.
• Current BAA methods rely on manual interpretation of hand radiographs, presenting a bottleneck in clinical practice.

Purpose of the Study:

• To develop and evaluate an automated computer vision system for BAA in children aged 0-18 years.
• To assess the efficacy of combining Histogram of Oriented Gradients (HOG), Local Binary Pattern (LBP), and Scale-Invariant Feature Transform (SIFT) for automated BAA.

Main Methods:

• Utilized 442 left-hand radiographs from the University of Southern California (USC) hand atlas.
• Applied HOG-LBP-SIFT features with background subtraction, extracting features from carpal and epiphyseal regions of interest (ROIs).
• Employed Support Vector Machine (SVM) with 5-fold cross-validation for classification.

Main Results:

• Achieved accuracies of 73.88% (female) and 68.63% (male) for BAA.
• Reported a mean absolute error of 0.5 years for both genders.
• Demonstrated high accuracy within 1-year (95.32% female, 96.51% male) and 2-year (100% female, 99.41% male) ranges.
• Cohen's kappa coefficients (0.71 female, 0.66 male) indicated substantial agreement with radiologist assessments (p < 0.05).

Conclusions:

• The proposed automated BAA method using combined HOG-LBP-SIFT features is robust, accurate, and easy to implement.
• This computer-aided diagnosis (CAD) approach significantly reduces processing time and the number of ROIs required for BAA.
• The system shows substantial agreement with expert radiologists, offering a viable tool for clinical use.