Skin lesion segmentation based on preprocessing, thresholding and neural networks

This abstract describes the segmentation system used to participate in the challenge ISIC 2017: Skin Lesion Analysis Towards Melanoma Detection. Several preprocessing techniques have been tested for three color representations (RGB, YCbCr and HSV) of 392 images. Results have been used to choose the better preprocessing for each channel. In each case a neural network is trained to predict the Jaccard Index based on object characteristics. The system includes black frames and reference circle detection algorithms but no special treatment is done for hair removal. Segmentation is performed in two steps first the best channel to be segmented is chosen by selecting the best neural network output. If this output does not predict a Jaccard Index over 0.5 a more aggressive preprocessing is performed using open and close morphological operations and the segmentation of the channel that obtains the best output from the neural networks is selected as the lesion. Image formats and preprocessing techniques Three image formats have been used: RGB, YCbCr and HSV. Each channel was treated separately. Preprocessing consisted on one or several of these techniques: • Anisotropic diffusion • Contrast enhancement • Color consistency • Gamma compensation • Color normalization 155 tests were performed on 392 images extracted from ISIC 2016 database. These images didn’t present hairs, reference circles or black frames. Jaccard Index was obtained and analyzed to choose the better thresholding segmentation for each channel. Conclusions are shown in Table 1. The decision was based on the following criteria: • The less preprocessing the better • The highest number of Jaccard Index above 0.8 • The best mean Jaccard Index of the segmentation of the 392 images • Number of best segmentations Two different preprocessing techniques were applied to the Blue channel because one of them was the segmentation that presented the highest number of best segmentations among all the experiments and the other had the best mean value for the experiments carried out on the blue channel. An Otsu thresholding was applied to each channel and the object that was closer to the center of the image was selected as the lesion. This selection was then compared with the ground truth and the Jaccard Index was calculated. As an example, some results are shown in Figure 1 and Figure 2. Table 1. Preprocessing applied to the different color channels Channel Preprocessing with the best results in terms of Jaccard Index R Color normalization and contrast enhancement G Gamma compensation, color consistency, conversion to HSV, anisotropic diffusion, contrast enhancement, conversion to RGB B Contrast enhancement Y Gamma compensation, color consistency, conversion to HSV, anisotropic diffusion, contrast enhancement, conversion to RGB, conversion to gray scale Cb Anisotropic diffusion Cr Anisotropic diffusion H Anisotropic diffusion and contrast enhancement S Gamma compensation, color consistency, conversion to HSV, contrast enhancement V Gamma compensation, color consistency, conversion to HSV B1 Gamma compensation, color consistency, conversion to HSV, anisotropic diffusion, contrast enhancement, conversion to RGB Figure 1. Results of experiments carried out on the blue channel applying different preprocessing techniques before thresholding. The experiments are sort from higher to lower mean Jaccard Index. Experiment 138 uses gamma compensation, color consistency, anisotropic diffusion and contrast enhancement. Figure 2. Number of best segmentations in each experiment. Experiment 89 uses contrast enhancement.