Bayesian variable selection for pre-cancerous versus cancerous tissue diagnosis using elastic scattering spectra. Jiao, Y., Diethe, T., Austwick, M., Lovat, L., Hopper, C., Bown, S., & Barber, D. In Biomedical Applications of Light Scattering IV, Proceedings of SPIE, volume 7573, 2010.
Bayesian variable selection for pre-cancerous versus cancerous tissue diagnosis using elastic scattering spectra [link]Paper  abstract   bibtex   
The aim of this study is to improve on the current diagnostic accuracy of elastic scattering spectroscopy by adopting an approach based on Bayesian linear methods, in particular Bayesian variable selection methods in logistic regression. The method attempts to automatically identify a small number of wavelengths that are most informative of the cancerous/pre-cancerous class label. This is potentially useful since rather than using broadband spectrum, a small groups of selected wavelengths bands may be used without significant loss of diagnostic accuracy, which will simplify the data collection by using photon detector. We also tried the approach which penalizing large local changes on the weight factors which will potentially help understanding the physical correlation and interpretations about components of the signals which are important for diagnosis. We apply the method to three independent dataset (High Grade Dysplasia in Barrett's Oesophagus (BE), metastastic breast cancer in lymph nodes, and malignancy in skin lesions) and compare the diagnosis accuracy with conventional SVM and LDA. The technique selects around 15 critical wavelengths, without significant degradation in test performance.For example, for Barrett's Oesophagus, based on a balanced test dataset of equal numbers of pre-cancerous and cancerous samples, the method predicts with an accuracy of 74% (This would give a specificity 87% and sensitivity 60%), competitive with the conventional state-of-the-art techniques.
@inproceedings{jiao2010bayesian,
	Abstract = {The aim of this study is to improve on the current diagnostic accuracy of elastic scattering spectroscopy by adopting an approach based on Bayesian linear methods, in particular Bayesian variable selection methods in logistic regression. The method attempts to automatically identify a small number of wavelengths that are most informative of the cancerous/pre-cancerous class label. This is potentially useful since rather than using broadband spectrum, a small groups of selected wavelengths bands may be used without significant loss of diagnostic accuracy, which will simplify the data collection by using photon detector. We also tried the approach which penalizing large local changes on the weight factors which will potentially help understanding the physical correlation and interpretations about components of the signals which are important for diagnosis.

We apply the method to three independent dataset (High Grade Dysplasia in Barrett's Oesophagus (BE), metastastic breast cancer in lymph nodes, and malignancy in skin lesions) and compare the diagnosis accuracy with conventional SVM and LDA. The technique selects around 15 critical wavelengths, without significant degradation in test performance.For example, for Barrett's Oesophagus, based on a balanced test dataset of equal numbers of pre-cancerous and cancerous samples, the method predicts with an accuracy of 74% (This would give a specificity 87% and sensitivity 60%), competitive with the conventional state-of-the-art techniques.},
	Author = {Yan Jiao and Tom Diethe and Martin Austwick and Laurence Lovat and Colin Hopper and Stephen Bown and David Barber},
	Booktitle = {Biomedical Applications of Light Scattering IV, Proceedings of SPIE},
	Date-Added = {2011-07-29 12:28:26 +0100},
	Date-Modified = {2015-10-28 13:59:14 +0000},
	Title = {Bayesian variable selection for pre-cancerous versus cancerous tissue diagnosis using elastic scattering spectra},
	Url = {http://www.tomdiethe.com/research/papers/8},
	Volume = {7573},
	Year = {2010},
	Bdsk-Url-1 = {http://www.tomdiethe.com/research/papers/8}}

Downloads: 0