The emerging paradigms in cancer research indicate the need for a multi-perspective and multimodal screening approach for early lung cancer diagnosis to increase the probability of clinical resection. Currently, no standalone screening methodology has been proven to suffice any clinical diagnosis. Computed tomography(CT) has been proved to present abnormality at an early stage with less impact on survival rate in population studies. Nevertheless, because of its non-invasive characteristic, it can be used for diagnosis, prognosis and visualization of tumor. Studies have shown that Computer aided diagnosis (CAD) as a second reader can perform in a similar capacity as a radiologist. The sensitivity and specificity can further be improved if CAD based CT is combined with content-based image retrieval (CBIR), where display of similar diagnostic proven cases can speed up the radiological analysis and also increase the effectiveness of the radiologist. Both the classification and the retrieval tasks have much to do with the human visual system (HVS). Objectiveness does not exist in the ability to detect and diagnose cancerous tissue on the CT by the HVS, nevertheless the CAD which is based on a computer vision system (CVS), can only perform as well as the HVS. The proposed approach for classification and retrieval relies on the mapping between the HVS and a CVS.;The segmentation of lung nodule is a prerequisite for both the CAD and CBIR tasks. A novel segmentation method is proposed which exploits the time map relationship between the hessian and level set based segmentations. The mapping is generated using the statistics from the hessian segmentation through a weighted regression model trained a priori. It is shown that the proposed computer based segmentation can perform as efficiently as the visual description of the radiologist to aid the retrieval type of tasks. The method exploits the intensity invariant properties of the eigenvalues from the hessian decomposition and the time crossing map from the level set approach to accurately determine the nodule boundary.;The classification part demonstrates that, for optimum selection of features, each feature should be analyzed individually and collectively with other features to evaluate the impact on the CAD system based on the class representation. This methodology will ultimately aid in improving the generalization capability of the classification module for early lung cancer diagnosis. Nonparametric correlation coefficients, multiple regression analysis and principle component analysis (PCA) were used to map the relationship between the represented features from the 4 radiologists and the computed features. Artificial neural network (ANN) is used for classification of benign and malignant nodules to test the hypotheses obtained from the mapping analysis.;The final part of the dissertation work includes a lung nodule based similar volume retrieval approach based on the signature generated from the selection in the high-level feature space. The signature is generated by representing the psychophysical similarity between low-level (content) and high-level (semantic) features as a Max-flow/Min-cut graph cut solution. The quantification of the similarity is done using a non-parametric rank correlation coefficient. The retrieval works on a hierarchical framework to emulate the clinical diagnosis processes. The selection and weightage of content features is automatically generated thus providing the necessary abstraction to the radiologist. The retrieval accuracy of the proposed approach is done in content domain for the five models generated in the semantic domain.
展开▼
