Diseño de un prototipo electrónico identificador de melanomas y lunares benignos, utilizando técnicas de machine learning.

Abstract

The present research study examines the capacity of Artificial Intelligence and Neural Networks to classify and identify patterns used in the area of medical imaging. The potential benefit of combining human knowledge based on scientific evidence and the compilation of a specific database was investigated. The gaol was to reduce the time gap in relation to clinical diagnosis and the results of said analysis being issued, which in turn would strengthen the efficiency and likely effectiveness of treatment. The methodology that was used is based on the collection of 2750 images corresponding to the training stage with 1372 images of moles, 474 images of melanomas and 254 images of seborrheic keratosis. In the test and validation stage, 471 cases of moles were used, 147 cases of melanomas and 132 cases of seborrheic keratosis. This set of data was entered into the inputs of neural networks as part of the newest “Machine Learning” technique. Regarding the multiclass task and supervised machine learning, a precision of 90,66% was obtained (16% Mole errors, 4% Melanoma errors, 8% Seborrheic Keratosis errors, with a mean error of 9,3%). This imbalance varies according to the number of stages used for training, in this case 43 hours of training were necessary to obtain the ideal model. It is important to take into consideration that correct classification is directly proportional to the quality of the data set and the number of cases inputted for learning.