An Estimation on the Carpal Bone Extraction Using Intensity Normalization

• Main Authors: Ting-Jyun Peng, 彭莛鈞
• Other Authors: Shyr-Shen Yu
• Format: Others
• Language: zh-TW
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/49397831076295340924

碩士 === 國立中興大學 === 資訊科學與工程學系 === 102 === Assessing bone age by viewing hand radiographs plays a critical role in clinical pediatric endocrinology. According to the past researches, to assess bone ages by phalanx features among 0 to 7 years old may get the worse bone age assessment accuracy as the smaller ages goes. The clinical researches had indicated that among the ages from newborn to about 7 years old, the maturation of carpal bones appear in a specific order and separate from each other. And the features among the period that the carpal bones begin fusing in about 8 years old has been proven effective in a research of recent years. For these reasons, to assess bone age had demonstrated to be reliable under this range of ages. However, the process of extracting the carpal region of interest (CROI) from radiographs is usually coupled with such image quality problems like low contrast, non-uniformly or over-low exposure. Hence it is quite challengeable task to separate the pixels represent the bone tissue from the soft tissue in radiographs. This dissertation proposed an image intensity normalization method which applied the transform function, z transform in Statistics. By transforming each intensity value found in an image to z value respectively, all the radiographs from heterogeneous sources can be normalized into one consistent standard. The pixels represent the carpal bones in radiographs are segmented by intensity thresholding. With respect to balanced accuracy, the chosen criteria of segmentation accuracy in this dissertation, the experiment results reveal proposed method raises the performance of control group obviously and outperforms the image intensity normalization methods including linear normalization, histogram equalization and contrast limited histogram equalization, which are often used in medical image processing. In addition, the average balance accuracy of the samples normalized by proposed method and grouped by rounded bone age all reach about 80 percent.