Computational models for skin cancer diagnosis based on optical spectroscopy and imaging

This dissertation focuses on the development of computational models and algorithms based on optical spectroscopy and imaging to assist skin cancer diagnosis in different clinical scenarios. Optical spectroscopy and imaging provide a wealth of biochemical and morphological information concerning the functional status and disease state of tissue. Interpreting this information can be challenging due to the high dimensionality of the data both spatially and spectrally. Therefore, we rely on complex computational models to simplify and analyze this information. First, we introduce our physiological model, a computational Monte Carlo lookup table inverse model, which can extract physiological parameters from diffuse reflectance spectroscopy (DRS) data. We applied this model on a clinical DRS dataset. Our findings suggest that DRS can reveal physiologic characteristics of skin and this physiologic model offers greater flexibility for diagnosing skin cancer than a purely statistical analysis. Then, we proposed using our model to assess tumor margins of non-melanoma skin cancer. With two independent clinical datasets, we trained models including DRS data normalization, a physiological model, parameter selection, and logistic regression classifiers using one dataset, and our test results on a second dataset with high accuracy showed that DRS can be potentially used to map the tumor margin prior to surgery and monitor margins during the surgery on the surface of the skin. For melanoma detection, to reduce unnecessary biopsies while still accurately detecting melanoma lesions, we proposed using principal component analysis (PCA) and logistic regression models based on Raman spectroscopy for generating a “second opinion” for lesions being considered for biopsy. Our work is a significant step toward the application of Raman spectroscopy for melanoma detection in the clinic. Also, we developed a feature engineering based similarity assessment algorithm to help with the intra-patient evaluation of moles, which will be helpful for finding melanoma lesions which are different from other lesions on the patient and typically at higher risk for malignancy. We found that our algorithm agrees well with three dermatologists in terms of the similarity of moles, which showed the potential of our algorithm to benefit melanoma detection.