Open Collections

Improving cervical neoplasia diagnosis via novel in vivo imaging technologies and deep learning algorithms

University of British Columbia

Two directions are explored for improving the current cervical cancer diagnosis procedure. The first investigates the future deployment of in vivo confocal imaging in the clinic, for detecting precancerous tissues, and the second proposes an algorithm for automatic interpretation of histology images (acquired by light microscopy). We acquired i) confocal microscopy images of cervical biopsies taken from 50 patients, at different tissue depths and ii) histology images of different sections cut from each biopsy. From the confocal images, we identified four features that carry enough information relevant to cell morphology and tissue architecture. We demonstrated that the relevant information in these features is comparable to that extracted from the same features in histology images. This implies that we can obtain the relevant information from confocal imaging, without having to cut a biopsy from the patient’s cervix. We then studied the confocal images and determined the grade lesion of every biopsy and found that confocal imaging resulted in less false positives than the diagnosis given by the gynecologist (based on the appearance of the cervix under colposcopy). Utilizing confocal microscopy technology in the clinic would thus decrease the number of unnecessary biopsies. We then developed a deep learning algorithm that automatically and quantitatively assesses HPV contaminated and proliferating cells in histology images of biopsy sections. The automatic assessment of this procedure is important as it plays a significant role in differentiating between disease grades but forms a challenging and complex task and demands a large amount of time when performed manually by a pathologist. We demonstrated that this algorithm could help the pathologists to differentiate between different grades of cervical precancerous tissues. Our results are also more reproducible compared to other methods (like color deconvolution) that are widely being used in the field of digital pathology. The in vivo imaging and automatic image analysis algorithms demonstrated in the thesis can potentially enable i) real time diagnosis in the clinic, and ii) fast interpretation of histology images in a reproducible and cost-effective manner. While developed for cervical neoplasia, these methods could be extended to oral cavity, skin, and other epithelial tissue cancers.

Text

2018-04-27

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/65676

Biomedical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/