Utilização de Redes Neurais Artificiais para Determinação de Propriedades Ópticas a partir de Medidas de Refletância Difusa

Skin cancer is the most frequent type of cancer in Brazil, mainly due to the excess of exposure to the Sun radiation. This disease is associated to a uncontrolled cell growth that can invade tissues and organs, and spreading to different body parts. One of the potential tool for skin cancer diagnostics is the Spacial Frequency domain imaging (SFDI), which determinates the optical properties of tissues in image by illuminating them with varying spatial light patterns. Despite its diagnostic potential, the optimization of the signal acquisition and image processing is currently being investigated, such as computation time efficiency and accuracy of the computed data. In this work, we investigated the optimization of the SFDI technique by employing Artificial Neural Networks (RNA) for pattern recognition and determination the optical properties of tissues from diffuse reflectance data. Values of diffuse reflectance for light beams with different spatial frequencies in different tissues were generated from diffusion theory and used as input data for the RNA supervised training. The output data consisted of the corresponding optical properties of the tissues (absorption and scattering coefficients). Feedforward RNAs with different number of neurons were evaluted, employing a backpropagation algorithm and sigmoid activation function. Results showed that the accuracy of the RNA in determining optical properties from diffuse reflectance data increased with the number of neurons, although the mean squared error were smaller than 10-8, for RNA with number of neurons between 5 and 10. Theses results indicated that, due to its processing speed and accuracy, the use of RNA can be a valuable tool for optinization of the SFDI technique for skin cancer diagnostics.