Segmentation and Feature Extraction of Hand Radiographs for Bone Age Assessment

Abstract

The Bone Age is a fairly reliable measure of persons growth and maturation of skeleton. Bone age assessment (BAA) has been a standard procedure to predict the age of a person from hand radiograph. The difference between chronological age and bone age indicates the presence of endocrinological problems. The automated bone age assessment system (ABAA) based on Tanner and Whitehouse method (TW3) requires monitoring the growth of radius-ulna and short bones (phalanges) of a left hand. The Tanner and Whitehouse - 3 (TW3) method involves assigning scores to the bones of interest of left hand and assessing the age of a person using the aggregate of scores. Due to the complexity and involved processing time, it is difficult for pediatricians to use the TW3 method in clinics. Hence, automating the whole procedure avoids human error at the same time reduces the processing time. This thesis investigates design and development of ABAA system to predict bone age of a person from hand radiograph. Fully ABAA system consists of 5 main stages which are (1)Pre-processing, (2)ROI extraction, (3)Segmentation, (4)Feature extraction and (5)Classification. In this context, two fully automatic extraction methods are proposed for the region of interest of phalanges (PROI) and radius-ulna bones (RUROI) using the left-hand radiograph. Experimental results demonstrate that fully automated PROI and RUROI extraction methods are simple, accurate and fast because underlying mechanism is free from complex mathematical procedure. Segmentation of hand bones plays a vital role in process of ABAA, though it is a challenging task to segment bones from the soft tissue. The problem arises because of overlapping pixel intensities between the bone region and soft tissue region and also overlapping pixels between soft tissue region and background. Hence, there is a need for a robust segmentation technique. In this context, two segmentation techniques are proposed, one for extracted PROI images and other for RUROI images. Quantitative ivand qualitative results of proposed segmentation techniques are evaluated and compared with other state-of-the-art segmentation techniques. The segmentation accuracy achieved by proposed segmentation techniques is 94 percent and is 97 percent on PROI and RUROI extracted images respectively. Medical experts have also validated the qualitative results of proposed segmentation techniques. Experimental results reveal that proposed techniques provide higher segmentation accuracy as compared to the other state-of-the-art segmentation techniques. Further, the development of an ABAA system requires robust feature extraction and efficient classification methods. In this context, this thesis implements and analyzes three different feature extraction methods namely texture feature analysis, Histogram of Oriented Gradients (HOG) and Bag of Features (BoF) methods on segmented PROI images. Also, three different classifiers namely Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Random Forest classifier are implemented to evaluate the performance of extracted features. Further, experimental results of BoF with random forest classifier yields a mean error (Merr) of 0.58 years and root mean square error (RMSE) of 0.77 years for the bone age range of 0-18 years and outperforms other existing BAA methods. Finally, experimental results have also proved that the use of gender bias improves the classification performance. The best performance is obtained from ring, middle and index fingers. Hence, the proposed fully automated technique can be used for bone age assessment of a person with enhanced accuracy.