Automatic Classification of Melanoma Using Grab-Cut Segmentation & Convolutional Neural Network

Background: Skin diseases are common health complications around the world. One of the most unsafe types of skin cancer is melanoma. Detection of skin cancer in the early stage can reduce the mortality rate. Manually detecting skin cancer from dermoscopy images is time-consuming, difficult, expensive, and requires experience. Therefore computer-based systems are necessary to make an early finding of skin cancer to plan timely treatment for patients to increase their survival rates. In this work, we introduce a Gaussian filter for noise removal. Furthermore, we applied the Grabcut segmentation technique to segment the affected lesions. Then we applied the segmented image into Convolutional Neural Network (CNN) with three hidden layers, including a flat and dense layer. The proposed CNN model uses the Adam optimizer to deliver the best performance with an accuracy value of 97.50%. In this study, we used 3500 images containing 500 images of each class. The data is further split into training and validation data as 75% and 25%. Based on the proposed model described in Fig. 2, segmented images are of size (32,32,3) as an input to the CNN model, which contains three hidden layers. The idea is convoluted, using 3 × 3 filters of 256,1n28 and 64, respectively, on each hidden layer. This Paper used the Convolutional Neural Network based method proposed for seven types of melanoma classification. It aids Skin cancer classifications that should be detectable by both patients and clinicians. By the experimental and evaluation, it can be said the designed model can be considered as a benchmark for skin cancer classification. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024.