Evaluation of a portable multispectral fluorescence lifetime device for skin cancer diagnosis using machine learning strategies

• Ano de defesa: 2020
• Autor(a) principal: Romano, Renan Arnon
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Diagnóstico de câncer de pele; Fluorescence lifetime imaging; Imagem de tempo de vida de fluorescência; Machine learning; Skin cancer diagnosis
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/76/76134/tde-29092020-154453/

Skin cancer diagnosis is a great challenge nowadays since it has the major incidence rates among all cancer types. Non-melanoma skin cancer is the most incident even being less aggressive, whereas melanoma is one of the most lethal cancer types due to its high metastatic potential. Early diagnosis is a main goal in new techniques development since it is responsible not only by increasing life expectancy, as well as decreasing treatment morbidity. Visual clinical inspection and then possibly dermoscopy and biopsy is the most common diagnosis procedure, being not so efficiently managed at the public health system, and can be very imprecise if not performed by a well-trained skilled dermatologist. An alternative early diagnosis technique is the label-free fluorescence since it is portable, real-time, non-invasive, not subjective, and that does not rely on exogenous fluorophores addiction. Some of the main native fluorophores are related to morpho-functional and metabolic activities on cell\'s, thus, they can act as intrinsic optical biomarkers (OB), such as collagen, elastin, Nicotinamide Adenine Dinucleotide (NADH) and Flavin Adenine Dinucleotide (FAD). Since cancer causes mitochondrial dysfunction on cells, these molecules are powerful endogenous agents for diagnosis. In order to study the effect of cells metabolic activities in autofluorescence, healthy fibroblasts cells, squamous cell carcinoma (SCC) and melanoma cancer cells were used. By two-photon excitation and multispectral steady-state and lifetime fluorescence microscopy, it was shown that optical redox ratio (ORR) and the lifetime parameters decrease in cancer cells, indicating that they may be relevant OB to distinguish cells. These OB were used in a clinical study on 119 patients with malignant lesions (melanoma, SCC, and nodular and superficial basal cell carcinoma (nBCC and sBCC)), and benign lesions (intradermal nevus (IN) and pigmented seborrheic keratosis (pSK)). Using a portable multispectral fluorescence lifetime imaging (FLIm) device, the patient skin cancer was imaged using a 355 nm excitation wavelength, and the fluorescence emission signal was detected in three channels mostly targeting collagen and elastin (390 ± 20 nm), NADH (452 ± 22 nm) and FAD (> 496 nm) emission. Using the OB extracted features, random forest, linear, and partial least square discriminant analysis classifiers were built on a training set and tested on a test set of images. As a result, to distinguish between healthy and nBCC tissue, classifiers showed that lifetime parameters are as important as steady-state fluorescence parameters, and achieved sensitivity of 88%. To distinguish between clinical similar malignant and benign lesions the models were tuned to achieve 100% sensitivity both on pigmented and non-pigmented lesions, which is very similar with the dermatologists\' biopsy accuracy (87-94%). Two-way and five-way classification models achieved accuracies around 70% and 60%, whereas dermatologists accuracies is around 66% in this clinical diagnostic classification. In general, using the machine learning models, the FLIm technique was able to provide a fast and dermatologist-level skin cancer screening, which may contribute especially in areas where there is a lack of trained professionals.