Hosseini, Rahil (2012) Fuzzy based approach for modelling uncertainty in classification for a computer aided detection. (PhD thesis), Kingston University, .
Abstract
A computerized image analysis technology suffers from imperfection, imprecision and vagueness of the input data and its propagation to all individual components of the technology including image enhancement, segmentation and object classification. Furthermore, a computerized medical image analysis system (CMIAS) deals with another source of ambiguity that is inherent in the image-based practice of medicine and intuitive knowledge of experts .. Therefore, a CMIAS such as computer aided detection (CAD) technologies implicitly suffer from uncertainty and vagueness both from image analysis techniques and medical diagnosis. Although several technology-oriented studies have been reported for CAD, no attempt has been made to address, model and overcome these types of uncertainty in the design of the CAD. However, uncertainty issues directly affect the accuracy of the system. This study addresses the main sources of the uncertainty in a CAD system. While uncertainty outcomes are latent in the input of a classifier, the aim is to model them in the classification for a CAD application. For this, this research takes advantages of type-2 fuzzy logic (T2FL). Integrating a T2FL model for object classification in CAD architecture allows us to model uncertainty issues. For this, an automatic approach models uncertainty in training dataset using membership function of a type-2 fuzzy set. This approach was applied to the candidate nodule classification problem in a lung CAD application. The ROC (receiver operating characteristic) analysis of the classifier results (with an average accuracy 95% (area under the ROC curve) for nodule classification) reveals that the T2FL is more capable of capturing the uncertainty in the model and achieving better performance results compared to type-l fuzzy logic counterpart. Furthermore, the research introduces the idea of uncertain rule-based pattern classification in environments which exhibit a lack of expert knowledge and with an imperfect training dataset. An automatic approach for rule extraction is presented which takes advantages of genetic algorithm for learning rule set of an T2FL system from training samples. The proposed approach was applied to the popular Wisconsin breast cancer diagnosis (WBCD) database. Analysis of the performance results reveals that this approach is competitive with, the best results of other proposed fuzzy classification methods to date in terms of trade-off between accuracy and interpretability, with an average accuracy of 96.6 % for the breast cancer diagnosis problem. This study introduces the concept of uncertainty in a CAD application. This is a first attempt toward modelling uncertainty issues in classification component for a CAD. The main contribution is automatically modelling uncertainties using membership functions and a rule set of a type-2 fuzzy logic. The performance evaluation on two different CAD classification problems (1) nodule classification in a lung CAD and (2) the WBCD diagnosis problem using Mammography CAD reveals the superiority of the T2FLS classifier for managing high levels of uncertainty compared to the T1FLS counterpart and providing classification that is more accurate. This approach is significant from two major aspects (l) clinical view: by producing more accurate results for diagnosis problems which can save more human lives, (2) technical view: modelling uncertainties in the design of a classifier using automatically presented approach for IT2FLS membership and rules generation. This is critical for multi-dimensional classification problems with large number of inputs and lack of expert knowledge as is the case for most of medical diagnosis problems.
Actions (Repository Editors)
 |
Item Control Page |