Abstract:

Abstract In recent days, Melanoma is found to be the most unpredictable and fatal form of skin disease. But it is curable if detected at the rudimentary stage. In this paper, Cascaded Correlation Neural Network, a new method of automatic classification of the skin images is presented. CCNN is self-organizing networks which by itself trains and add on new hidden layers consecutively till the error is minimized. By adopting this particular feature, an accurate and efficient image processing technique is implemented in this paper for cancer detection. As a preprocessing step, noise is filtered, and contrast enhancement is done using histogram equalization method. The color attributes are taken from RGB and opponent color space in the skin lesion and are provided as input to the CCNN. The proposed approach is tested on the ISIC database of melanoma images. Receiver Operating Characteristic curve is used to detect the performance of the suggested system. In the results obtained with 91.1 % accuracy, the sensitivity is 91.7% and the specificity is 89.2%. The result shows the potential of the proposed CCNN network.

Keywords: