Open Collections

Machine learning for cancer detection, grading, and prognosis : automated segmentation, patterning recognition, and AL-based analysis

University of British Columbia

In this thesis, we presented the design steps for developing new, reliable, and cost-effective diagnostic and prognostic tools for cancer using advanced Machine Learning (ML) techniques. We proposed tools to improve the diagnostic, prognostic and detection accuracy of quantitative digital pathology by incorporating advanced image analysis, image processing, and classification methods. In this thesis we presented our ML image-based analytic approaches for three cancer types (prostate, cervix, and kidney) with different scale ranges from the sub-micron to multiple centimeters. In this thesis, we demonstrated the full workflow to design an automated prognostic and grading system specially designed for prostate cancer. We started with demonstrating techniques for prostate glandular structures detection. Next, we introduced an automated cell segmentation method along with an interactive segmentation correction method requiring minimum user-interaction and finally, we introduced our ML classification algorithms. We trained our ML method on the features extracted from cells/nuclei that were segmented via our proposed techniques. Next, we studied renal carcinoma. We presented the workflow of renal carcinoma classification from image processing to feature selection and development of machine learning classification techniques. We extracted the features from renal vessel structures and demonstrated the design steps of machine learning classifiers to discriminate between different renal carcinoma subtypes using these features. The last cancer site studied was the cervix. We applied our techniques for cervical pre-cancer abnormality detection. We showed the whole pipeline of designing an automated classification method starting from tissue imaging to the development of ML classifiers using both classical and deep-learning methods. Although we conducted these studies on specific cancer types, a modified version of our algorithm could be applied to other cancers and disease sites. These techniques have great potential to improve the healthcare environment by providing extra information/second opinion to the medical experts or to be used as a part of the first line of a screening program.

Text

2019-12-24

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/72960

Interdisciplinary Oncology

University of British Columbia

UBCV

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