A Probabilistic Framework for Detection of Skin Cancer by Raman Spectra

This Ph.D. thesis focuses on objective methods for diagnosing skin cancer from Raman spectra. A method for suppressing background noise and dimension reduction in Raman spectra is suggested. A robust Bayesian framework for training a neural network is proposed, including an overfit control and outlier framework. Finally a visualization scheme for extracting important features from the trained neural network classifier based on sensitivity analysis is defined.

The performance on two types of skin cancer showed that 97.9% of basal cell carcinoma were identified correctly and 85.5% of malignant melanoma. The neural network classifier visualization showed that frequency bands, previously identified by visual inspection of Raman spectra by medical experts, were considered important for classification. Moreover, frequency band not previously used for skin lesion classification were identified. These identified important features are shown to originate from molecular structure changes in lipids and proteins.

While the theme of this dissertation is skin cancer diagnosis from Raman spectra, the dimension reduction and the neural network classifier can be applied in general to other types of pattern recognition problems.