NON INTERACTIVE GRABCUT BASED SEGMENTATION WITH DEEP LEARNING MODEL FOR SKIN LESION DETECTION AND CLASSIFICATION

Skin lesion segmentation and classification are crucial processes due to the visual resemblance among the benign and melanoma lesions. Robust and effective skin lesion classification is a major task to attain proficient skin lesion diagnosis. Because of the unclear boundaries among the tumor and skin, the segmentation of skin lesion becomes a tedious task. This paper presents a new deep transfer learning (DTL) based segmentation and classification model for diagnosing skin lesions from dermoscopic images. The presented model initially performs bilateral filtering based preprocessing to discard the noise exist in the image. In addition, Noninteractive GrabCut Algorithm is applied to perform the image segmentation process. Afterward, DTL model based on VGGNet-19 Network is utilized to extract a useful set of feature vectors. At last, extreme gradient boosting (XGBoost) and Linear Discriminant Analysis (LDA) models are used to identify the different class labels of skin lesions. For validating the proficient diagnosis outcome of the proposed model, a series of simulations were performed to exhibit the supreme performance of the presented method. From the obtained set of simulations, it is apparent that the VGG19-XGBoost model has outperformed the compared methods with the maximum sensitivity of 95.08%, specificity of 99.17%, and accuracy of 98.65%.