| Summary: | 博士 === 國立中興大學 === 資訊科學與工程學系所 === 101 === In recent years, computer-aided diagnosis (CAD) techniques have become one of the major research subjects in medical fields. Through the assistance of a successful CAD technique, the quality and productivity of doctors’ tasks can be improved in the diagnostic process. Recently, the use of fractal geometry for medical image analysis has gained prominence. Therefore, this dissertation develops different CAD techniques based on fractal theory for the following medical applications.
For classification of pathological prostate images, this dissertation proposes two fractal dimension texture features that can be extracted through differential box-counting method and our entropy-based fractal dimension estimation method and then combines them together as a fractal-based feature set to analyze variations of intensity and texture complexity in pathological images of prostatic carcinoma. Each image can be classified into an appropriate grade by using Bayesian, k-NN, and support vector machine (SVM) classifiers, respectively. Experimental results show that the proposed feature set is better than the feature sets extracted from existing texture analysis methods, such as multiwavelets, Gabor filters, and gray-level co-occurrence matrix methods in terms of the size of the feature set and discriminating capability in grading prostate images.
For segmentation of pulmonary nodules, this dissertation presents a method based on multifractal analysis to detect and segment pulmonary nodules from thorax computed tomography (CT) scans automatically. The proposed method does not have the problem of seed selection such as in region-growing and is able to detect and segment lung nodules successfully.
Finally, a CAD technique is developed to distinguish between benign and malignant nodules on thin-section CT scans. By combining a perfusion CT feature with the texture feature set based on the fractional Brownian motion method, the proposed feature set is used to analyze the variations of CT number and texture roughness in the tumor region. Experiment results show that the proposed method can achieve satisfactory diagnostic performance. In addition, image processing time and the amount of radiation exposure to patients can be reduced in the diagnostic process, because the proposed method only needs single post-contrast scan after injection of the contrast medium.
|
|---|
