Feature description with SIFT, SURF, BRIEF, BRISK, or FREAK? A general question answered for bone age assessment

• General purpose algorithms can give good results for a specific clinical task.
• Two methods for keypoint selection were applied: sparse and dense feature points.
• SIFT, SURF, BRIEF, BRISK and FREAK were evaluated/compared for bone age assessment.
• BAA is performed by extracting features from radiographs with SVM classification.

Solving problems in medical image processing is either generic (being applicable to many problems) or specific (optimized for a certain task). For example, bone age assessment (BAA) on hand radiographs is a frequent but cumbersome task for radiologists. For this problem, many specific solutions have been proposed. However, general-purpose feature descriptors are used in many computer vision applications. Hence, the aim of this study is (i) to compare the five leading keypoint descriptors on BAA, and, in doing so, (ii) presenting a generic approach for a specific task. Two methods for keypoint selection were applied: sparse and dense feature points. For each type, SIFT, SURF, BRIEF, BRISK, and FREAK feature descriptors were extracted within the epiphyseal regions of interest (eROI). Classification was performed using a support vector machine. Reference data (1101 radiographs) of the University of Southern California was used for 5-fold cross-validation. The data was grouped into 30 classes representing the bone age range of 0–18 years. With a mean error of 0.605 years, dense SIFT gave best results and outperforms all published methods. The accuracy was 98.36% within the range of 2 years. Dense SIFT represents a generic method for a specific question.