Study On The Key Techniques Of Computer-aided Diagnosis For Lung Cancer In Medical Images

Recently, with the advent and popularity of digital medical equipments and the development of technology of computer application, computer-aided diagnosis (CAD) has become one of the major research subjects in medical imaging and computer science. CAD is a way in which the computer processes the digital medical image and provides some information with reference value for the clinical radiologists’ decision, and reduces the load and the omission when radiologists read individually.In this thesis, the subjects of this study were lung cancer in computed tomography (CT) images and multi-projection correlateion images, and the problems of this study were the key techniques in the development of CAD. We have done experiments with85lung CT images from Lung Image Database Consortium (LIDC) and59multi-projection correlateion images from Departmet of Radiology, Duke University. The experimental results indicated that our CAD schemes achieved higher performance levels for lung nodule detection than existing CAD schemes. The main work in this thesis included:(1) A lung nodule detection scheme in CT using a sequential uniform viewpoint reformation (SUVR) techniqueA key issue in current CAD schemes for nodule detection in CT is the large number of false positives, because current CAD schemes use only global3D information and discard useful local2D information. Thus, we developed a CAD scheme using the SUVR technique to obtain the2D local information. The SUVR technique first created sequential uniformly-distributed viewpoints around a nodule candidate, and decomposed a3D nodule candidate in3D image into a set of2D nodule candidates in a set of reformatted2D images generated from all possible viewpoints. Because a3D nodule would appear as circular objects in nearly all viewpoints, whereas a3D blood vessel almost always appeared as non-circular linear objects in some "effective" viewpoints, if we took advantage of these "effective" viewpoints, we can reliably remove many blood vessels. The CAD scheme based on the SUVR technique consisted of five steps for identification of initial nodule candidates, generation of2D reformatted images using the SUVR technique, nodule candidate segmentation in the2D reformatted images, feature determination, and false-positive reduction by use of a piecewise linear classifier. The experimental results indicated the CAD scheme based on the technique can markedly improve the performance for nodule detection.(2) High performance lung nodule detection schemes in CT using local and global informationFive high performance lung nodule detection schemes in CT were developed using local and global information, and the information was fused based on features, detection methods, and detection results. The fusion scheme based on features used2D and3D features simultaneously. The fusion scheme based on detection methods first used2D (3D) features, and then used3D (2D) features after the sensitivity of85%. The fusion scheme based on detection results used2D and3D features to detect nodules respectively, and the final result was the intersection or the union of these two detection results. The experimental results indicated the performance levels of these five schemes were similar, and they were much higher than the performance levels of using olye local or global information.(3) The relationship between the performances of nodule segmentation and detectionTwo assumed conclusions were acquired based on the theoretical analysis about the relationship between the performances of nodule segmentation and detection:under the given conditions, the relationship between the performances of nodule segmentation and detection is unconspicuous using different segmentation schemes, and the relationship is approximate linear using the same segmentation scheme with different parameters. Then, we did experiments using the empirical threshold scheme, OTSU threshold scheme, fuzzy C means scheme, Gaussian mixture models scheme, CV model scheme, and the region and edge-based active contour model scheme. The experimental results proved that the two assumed conclusions were correct. These two conclusions had profund and directive infulence on the study of CAD schemes of nodule detection in the further.(4) Lung nodule detection schemes based on multi-projection correlation imagesA conventional CAD scheme and a fusion CAD scheme based on the multi-projection correlation images were developed in this thesis. The conventional CAD scheme consisted of four steps for identification of initial nodule candidate in lungs, nodule segmentation based on dynamic programming, feature determination, and false positive reduction by use of a stepwise linear classifier. It processed each of three images of a subject independently and discarded the correlation information between the three images. The fusion CAD consisted of the four steps in the conventional CAD and two additional steps for registration of all candidates in the three images of a subject and integration of correlation information between the registered candidates. Because a nodule was more possible to be detected in multiply images of a subject than a false positive, many false positives would be removed by removing those candidates detected only once in the three images. The experimental results indicated that compared with the conventional CAD scheme, the fusion CAD scheme could greatly improve the performance for lung nodule detection.