| Summary: | 碩士 === 臺灣大學 === 醫學工程學研究所 === 98 === Statistics from health department shows that the lung cancer has been on the second place for male and the first place for female in causes of death by malignant tumors for more than twenty consecutive years. Since the survival rate of late stage
lung cancer is relatively low, early diagnosis is critical to survival.
Low dose CT is capable of detecting lung tumors of length 1-2 mm in diameter, however the CT interpreting of which remains as a non-trivial task. For example, to identify round-shape tumor and tube-shape blood vessel structure, a physician must undergo a time-consuming and error-prone process to examine each CT carefully. And it is even more difficult when it comes to find small nodules in this way. A physician must go beyond naked-eyes examination for better precision and efficiency.
Computer aid has been proven to be able to assist physicians in improving detection rate of lung tumors in CT scans, and it is getting wider popularity in the business of medical instruments. However, detection software in either commercial products or laboratory developments has yet much room to improve. Therefore it is an essential subject of research to improve the performance of computer aided detection
software.
This research goes through a series of analysis on pulmonary nodules and presents a set of methods for computers to be able to learn identifying tumors, and it is hoped that these methods may serve as diagnosis aids for doctors in the future. NATIONAL CANCER INSTITUTE (NCI) provides essential data including computed tomography (CT) images of around 397 patients with doctors'' marking.
The first step of the research is to automatically identify nodules by computer and store the data of nodules and nearby lung wall and blood vessels. The second step is to automatically partition the data by the types of nodules and perform a series of characteristic analysis. The third step is to train a automatic operator that classifies tumors on the basis of the characteristic data obtained in previous step, and the trained operator is put to work identifying nodules and then be evaluated for performance.
With the tumor data of 397 patients as the basis of training, the automatic operator achieves a satisfactory performance. The classifier trained out of the samples can achieve 83% up correction rate on average on the selected samples.
|
|---|
