Please use this identifier to cite or link to this item:
|Title:||Robust classification and detection with applications in biomedical images||Authors:||Lin, Dongyun||Keywords:||DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision||Issue Date:||2018||Source:||Lin, D. (2018). Robust classification and detection with applications in biomedical images. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||In this thesis, robust classification and detection methods with applications in biomedical images have been studied. The study is motivated by Chow’s theoretical works on classification with a reject option and specific applications in biomedical images. Technically, the study has been carried out from two perspectives. The first is proposing robust classification methods with a reject option based on Chow’s theoretical works to select ambiguous and outlier samples. The second is proposing robust multi-stage methods for biomedical image classification and detection. Firstly, we propose a twin support vector machine with a reject option (RO-TWSVM) through the Receiver Operating Characteristic (ROC) curve for binary classification. The method is formulated under a cost-sensitive framework and follows the principle of minimization of the expected classification cost. Extensive experiments are conducted on synthetic and real-world datasets to compare the proposed RO-TWSVM with the original TWSVM without a reject option (TWSVM-without-RO) and the existing SVM with a reject option (RO-SVM). The experimental results demonstrate that our RO-TWSVM significantly outperforms TWSVM-without-RO, and in general, performs better than RO-SVM. Secondly, we propose a novel two-stage method for robust biomedical image classification. This method is proposed based on a cascade of a support vector machine (SVM) with a reject option and subspace analysis to robustly classify biomedical images. At stage 1, we propose a confidence score based classification rule with a reject option for a preliminary decision using the SVM. The testing images going through stage 1 are separated into two groups based on their confidence scores. Those testing images with sufficiently high confidence scores are classified at stage 1 while the others with low confidence scores are rejected and fed to stage 2. At stage 2, the rejected images from stage 1 are first processed by a subspace analysis technique called eigenfeature regularization and extraction (ERE), and then classified by another SVM trained in the transformed subspace learned by ERE. Extensive experiments implemented on three benchmark datasets show that the proposed method significantly outperforms several state-of-the-art competing methods in terms of multiple classification performance metrics. Thirdly, we propose a novel two-stage method for robust biomedical image detection. We investigate a specific task of detecting ring-like endosomes in fluorescent microscopy images. The proposed method first selects candidate patches using local feature information. Then, it identifies endosomes from candidate patches using a pretrained SVM based on global information of training patches. The experiments conducted on real-world microscopy images show that the proposed method significantly outperforms several state-of-the-art competing methods in terms of multiple detection performance metrics. Overall, the thesis contributes to two types of novel methods to enhance the robustness of classification and detection methods: (i) Proposing a twin support vector machine with a reject option (RO-TWSVM) through the Receiver Operating Characteristic (ROC) curve to select ambiguous and outlier testing samples; (ii) Proposing multi-stage methods for the classification or detection of biomedical images.||URI:||https://hdl.handle.net/10356/90200
|DOI:||10.32657/10220/47324||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.