COMPUTATIONAL TECHNIQUES FOR SKIN LESION TRACKING AND CLASSIFICATION

Resource type
Thesis type
(Thesis) Ph.D.
Date created
2014-11-17
Authors/Contributors
Abstract
We propose image-based automatic pigmented skin lesion (PSL) tracking and classification systems for early skin cancer detection. The input to our PSL tracking system is a pair of skin back images of the same subject. The output is the correspondence (matching) between the detected lesions and the identification of newly appearing (or disappearing) ones. We start by automatically detecting a set of anatomical landmarks by globally optimizing a pictorial structure. The detected landmarks are used to restrict the search space during lesion localization and encode the anatomical spatial context of lesions using a set of Jacobian based features, which are useful for lesion matching. The matching step is performed by an uncertainty-based feature learning approach using a high order Markov Random Field (MRF) optimization framework. The LND detection and PSL matching steps involve the optimization of energy functions with hyper-parameters, which are learned using a structured support vector machine. Given the dependence of the lesion matching on the detected landmarks, we propose an adaptive system that predicts the landmark detection error and leverages it to automatically adapt the lesion matching objective function. We also make the following contributions in our PSL classification system. In our work, we focus on extracting features for streak detection due to the clinical importance of the absence or presence of the streaks in dermoscopic images. To this end, we develop a novel hair disocclusion method using dual-channel quaternion tubularness filters and MRF-based multi-label optimization. To facilitate a comprehensive evaluation on hair segmentation, we provide a publicly available new hair simulator software. Further, integrating the quaternion tubularness filters and context of the eigenvectors, we propose a novel lesion descriptor. At the end, we apply weakly supervised learning approaches to perform the PSL classification task.
Document
Identifier
etd8810
Copyright statement
Copyright is held by the author.
Permissions
The author has not granted permission for the file to be printed nor for the text to be copied and pasted. If you would like a printable copy of this thesis, please contact summit-permissions@sfu.ca.
Scholarly level
Supervisor or Senior Supervisor
Thesis advisor: Hamarneh, Ghassan
Thesis advisor: Lee, Tim K.
Member of collection
Attachment Size
etd8810_HMirzaalianDastjerdi.pdf 14.53 MB