2v-SVM-based semi-supervised learning with application to carotid plaque characterization

Accumulated plaque in the carotid artery can be either stable or unstable. In the latter case, pieces from the plaque may break off and travel to the brain, potentially blocking blood flow and causing a stroke. The assessment of plaque stability is usually done visually based on its ultrasound image. Recently, Computer Aided Diagnostics (CAD) tools have been proposed to support clinical decisions that use machine learning techniques to analyze carotid plaque ultrasound images, and classify plaques as stable or unstable. Previous work has used supervised or unsupervised learning methods for this task, with the Support Vector Machine (SVM) being one of the most widely used supervised classifiers. However, only limited labelled data is available for training since the "gold standard" for plaque characterization as stable or unstable is to perform biopsy after endarterectomy. On the other hand, unlabelled data are readily available. A further complication arises from the fact that there is an imbalance in class labels (in our case stable or unstable), with data labelled as unstable dominating the training data set. The objective of this thesis is to investigate whether the use of unlabelled data in a semi-supervised algorithm can lead to better classification accuracy than that from a purely supervised classifier. To that end, we propose to use the 2ν-SVM classifier (a version of the SVM which efficiently handles cases with imbalance in the class labels) in the Branch-and-Bound (BB) implementation of a semi-supervised SVM. We investigate the performance of the proposed method on both synthetic data, specifically the popular 2-Moon data set, and on carotid plaque ultrasound images, where amongst the set of labelled plaque images there is at least twice as much unstable plaques than stable plaques. For the 2-Moon data with class imbalance in the labelled data set, the use of 2ν-SVM with a semi-supervised algorithm yielded an improved classification result over the same semi-supervised algorithm in combination with the generic SVM, and also over pure supervised SVM classifiers. In the case of carotid plaque images, for which the training data was composed of 18 labelled (6 stable, 12 unstable) and 18 unlabelled images - the results were inconclusive. A highest classification accuracy of 70.11% was achieved with both a supervised 2ν-SVM, and the BB (semi-supervised) algorithm that used the 2ν-SVM as a classifying component. The plain SVM, either when used as a purely supervised classifier, or as part of the BB algorithm, yielded lower classification accuracy, proving that the 2ν-SVM better handles cases when the labelled training data contains imbalanced numbers of class labels.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.