Desenvolvimento de software e hardware para diagnóstico e acompanhamento de lesões dermatológicas suspeitas para câncer de pele

Abstract

Cancer is responsible for about 7 million deaths annually worldwide. It is estimated that 25% of all cancers are skin, and in Brazil the most frequent in all geographic regions type. Among them, the melanoma type, accounting for 4% of skin cancers, whose incidence has doubled worldwide in the past decade. Among the diagnostic methods employed, it is cited ABCD rule which considers asymmetry (A), edges (B), color (C) and diameter (D) stains or nevi. The digital image processing has shown good potential to aid in early diagnosis of melanoma. In this sense, the objective of this study was to develop software in MATLAB® platform, associated with hardware to standardize image acquisition aiming at performing the diagnosis and monitoring of suspected malignancy (melanoma) skin lesions. Was used as the ABCD rule for guiding the development of methods of computational analysis. We used MATLAB as a programming environment for the development of software for digital image processing. The images used were acquired two banks pictures free access. Images of melanomas (n = 15) and pictures nevi (not cancer) (n = 15) were included. We used the image in RGB color channel, which were converted to grayscale, application of 8x8 median filter and approximation technique for 3x3 neighborhood. After we preceded binarization and reversing black and white for subsequent feature extraction contours of the lesion. For the standardized image acquisition was developed a prototype hardware, which was not used in this study (that used with enclosed diagnostic images of image banks), but has been validated for evaluation of lesion diameter (D). We used descriptive statistics where the groups were subjected to non-parametric test for two independent samples Mann-Whitney U test yet, to evaluate the sensitivity (SE) and specificity (SP) of each variable, we used the ROC curve. The classifier used was an artificial neural network with radial basis function, obtaining diagnostic accuracy for melanoma images and 100% for images not cancer of 90.9%. Thus, the overall diagnostic accuracy for prediction was 95.5%. Regarding the SE and SP of the proposed method, obtained an area under the ROC curve of 0.967, which suggests an excellent diagnostic ability to predict, especially with low costs, since the software can be run in most systems operational use today.