Deep learning for lung cancer analysis

Anderson, Owen (2023) Deep learning for lung cancer analysis. EngD thesis, University of Glasgow.

Abstract

This thesis describes the development and evaluation of two novel deep learning applications that tackle two cancers that affect the lungs. The first, lung cancer, is the largest cause of cancer-related deaths in the United Kingdom. It accounts for more than 1 in 5 cancer deaths; around 35,000 people every year. Lung cancer is curable providing it is detected very early. Computed tomography (CT) X-ray imaging has been shown to be effective for screening. However, the resulting 3D medical images are laborious for humans to read, and widespread adoption would put a huge strain on already stretched radiology departments. I developed a novel deep learning based approach for the automatic detection of lung nodules, potential early lung cancer, that has potential to reduce human workloads. It was evaluated on two independent datasets, and achieves performance competitive with published state-of-the-art tools, with average sensitivity of 84% to 92% at 8 false positives per scan. I developed a related invention which allows hierarchical relationships to be leveraged to improve the performance of CAD tools like this for detection and segmentation tasks.

The second cancer is malignant pleural mesothelioma. It is very different from lung cancer: rather than growing within the lung, mesothelioma grows around the outside of the lung in the pleural cavity, like the rind on an orange. It is a rare cancer, caused by exposure to asbestos fibres. It can take decades from exposure to symptoms developing. In Glasgow many mesothelioma patients worked in the ship-building industry, which relied heavily on asbestos. Although asbestos has been banned in the UK since 1999, its presence in buildings and equipment built before then mean that mesothelioma will remain a problem for years to come. Sadly, asbestos is still being mined and many countries, including the United States, have still not instigated a complete ban. For mesothelioma the main challenge is not detection, but accurate measurement —- without the ability to measure tumour size it is difficult to evaluate potential treatments. We therefore developed a fully automated volumetric assessment of malignant pleural mesothelioma. Performance of the algorithm is shown on a multi-centre test set, where volumetric predictions are highly correlated with an expert annotator (r=0.851, p<0.0001). Region overlap scores between the automated method and an expert annotator exceed those for inter-annotator agreement across a subset of cases. Dice overlap scores of 0.64 and 0.55, by cross-validation and independent testing respectively, were achieved. Future work will progress this algorithm towards clinical deployment for the automated assessment of longitudinal tumour development.

• Item Type: Thesis (EngD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; Q Science > QA Mathematics > QA76 Computer software; T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Siebert, Dr. Paul, Lavery, Professor Martin and Goatman, Dr. Keith
• Date of Award: 2023
• Depositing User: Theses Team
• Unique ID: glathesis:2023-83850
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 24 Oct 2023 09:57
• Last Modified: 24 Oct 2023 09:58
• Thesis DOI: 10.5525/gla.thesis.83850
• URI: https://theses.gla.ac.uk/id/eprint/83850

Related URLs

• Article DOI
• Article DOI
• Article DOI
• Article DOI