Fourier Transform infrared spectroscopy, in particular, infrared microspectroscopy, has great potential for clinical applications in the flow of cancer diagnosis. Using large focal plane array detectors and with advancements in computer power, infrared hyperspectral imaging has significant advantages in both accuracy and speed of diagnosis. In light of previous research on cancer diagnosis and digital histopathology using infrared imaging, further studies combined with machine learning algorithms have been conducted and are presented in this thesis. Human tissue samples including breast and prostate have been studied. Initial studies have been conducted on breast tissue on CaF2. Infrared images were obtained and analysed using two machine learning algorithms namely Random Forest and AdaBoost. This demonstrated that good classification results, classification accuracies of 89% and 92%, could be obtained to distinguish cancerous from normal associated tissue (NAT). The caveolin-1 stain was applied as a possible breast cancer diagnosis correlated stain. Classification accuracies on cancerous and NAT spectra were 100% and 71.4% respectively in the independent test, which indicates the great potential of caveolin-1 as a biomarker correlated with breast cancer diagnosis. For further implementation of infrared spectroscopy into clinical field, glass substrates, which are cheap and robust, are selected as potential new substrate for infrared disease diagnosis. Studies related to the performance of cancer diagnosis and digital H&E staining using infrared spectra collected from glass slides were conducted on breast tissue. Excellent separation between cancerous and NAT spectra was obtained with classification accuracies of 81.3% and 83.2% on cancer and NAT classes in the independent test. In addition, unbalanced classes are commonly observed in breast tissue analysis, as the epithelium cells are often much fewer in number compared with the stroma cells. A study using different sampling methods and classification methods to solve the problem and boost the classification results was conducted on the spectra collected from breast tissue on CaF2. Lastly, to test whether similar performance of classification can be observed from other types of tissue, studies on prostate tissue with glass substrates were also conducted. Reasonable classification results, classification accuracies, 72% and 68% were obtained with threshold (85% top scored testing spectra) added in the independent test (10 cores).
Date of Award | 1 Aug 2020 |
---|
Original language | English |
---|
Awarding Institution | - The University of Manchester
|
---|
Supervisor | Nicholas Lockyer (Supervisor) & Peter Gardner (Supervisor) |
---|
Developments in infrared spectral histopathology using machine learning algorithms
Tang, J. (Author). 1 Aug 2020
Student thesis: Phd