AbstractOptical projection tomography enables 3-D imaging of colorectal polyps at resolutions of 5 – 10 μm. This thesis presents image analysis methods for the polyp diagnosis from such images. Specifically, we investigate 3-D texture-based recognition methods, as well as weakly supervised classification methods, for the diagnostic task of discriminating levels of dysplastic change.
Firstly, we build a patch-based recognition system and evaluate both multi-class classification and ordinal regression formulations. 3-D texture representations computed with a hand-crafted feature extractor, random projection, and unsupervised image filter learning are compared using a bag-of-words framework.
Secondly, two novel classification methods are proposed to learn from partially and weakly annotated images respectively. For the partially annotated images, we developed a relevance ranking method to infer the overall classification model using unlabelled contextual image patches. For the weakly annotated images labelled at the image level, we proposed a boosting with regularised tree algorithm to learn the region classifier.
Results on a database of 90 polyps demonstrate that randomly projected features are effective. Discrimination was improved by carefully manipulating various important aspects of the system, including class balancing, output calibration and approximation of non-linear kernels. For the cancer region classification measured by the area under the ROC curve, 0.81 was achieved by training with image level labels, 0.85 by training with eight mouse click annotations per image. They both outperformed the competing methods. 0.88 was achieved by training with the delineated regions.
|Date of Award||2015|
|Supervisor||Chunhui Li (Supervisor) & Stephen McKenna (Supervisor)|