Computer-Aided Detection Based On Textural Features And High-performance Classifiers

Computer-aided detection (CAD) can help doctors interpret mass images and get valuable diagnostic information with digital image processing technique, computer visualization and so on. CAD system not only helps improving the sensitivity and specificity of diagnosis, and reducing doctor's interpreting time and workload, but also alleviates inconsistence in image interpretation due to different experiences or criteria used by radiologists.With the clinical application of virtual endoscopy (VE), CAD has demonstrated great potential with higher sensitivity and specificity. Most CAD systems for VE have been developed based on geometrical features, which can reflect morphological changes but be incapable of differentiating suspicious areas with similar shapes, such as distinguishing colonic polyps from folds or feces with similar shapes. Additionally, detection based on geometrical features is hard to obtain some critical features or diagnostic information because of failing to reflect diseased and normal tissues density changes, such as the invasion depth of bladder tumor into adjacent wall. The study indicates that texture can reflect the difference between diseased and normal tissues effectively. The effective combination between textural features and the high-performance classifier is expected to improve the performance of CAD system for detecting the colonic polyps with lower false positive rate and the invasion depth of bladder tumor. In view of the above questions, this study is mainly on the following research:1) CAD for bladder tumor and its invasion depth based on MRI imagesBladder tumor is a severe common tumor with high recurrence rate. The virtual cystoscopyis based on CT/MRI mostly and realizes the internal structure of 3D imaging, but it can't detect effectively the bladder tumor and its invasion depth into adjacent wall. The study indicates that the textures between bladder tumor and the wall are different in the images. Firstly, this study analyzes and screens some effective textural features. Then, the features are extracted based on a cell which is a small region. Finally, combining the improved classifier and the cell labeling method proposed, the CAD system for bladder tumor and its invasion depth is built. Experiment results using 16 patients'MRI datasets show42 out of the 56 images are correct, indicating that the sensitivity of the proposed detection scheme is about 82.2%, while labeling results of 15 out of 16 patients'consistented well with histopathological tests, with an accuracy of 93.8%. With texture-based classification, the CAD implemented preliminary detection of bladder tumor and its invasion depth, which might be a novel and promising tool for noninvasive tumor detection with virtual cystoscopy.2) CAD for colon polyps based on 3D textural featuresColorectal tumor incidence and mortality are both the third in malignant tumors, with a rising trend in China.A non-invasive detecting and screening method is needed urgently. Nowadays, virtual colonoscopy is based on geometrical features mostly, and has a high false positive rate because of the influences of folds and feces. Based on the textural differences between polyps and other tissues, the improved CAD system combines the textural features extracted from 3D region of interest (eROI) and the improved classifier, analyzes and classifies the different textural feature vectors. Experiment results using 13 patients'CT datasets from NIH show that the combination of the vector with 22 normal distribution features and the improved SVM classifier is best, and the specificity is 89.3% when the sensitivity is 100%. That indicates the ascendant of 3D textural features in detection of the colon polyps.