Skin Lesion Semantic Segmentation Using Convolutional Encoder Decoder Architecture

Computerized skin lesion analysis system often used segmentation technique which is always advantageous due skin lesion unequal size, shape and border. In this research paper deep convolutional encoder decoder neural network is proposed for pixel-wise semantic segmentation of dermascopic image of skin lesion. Proposed segmentation network consist sequence of encoder block and subsequent decoder block and final output is fed to pixel-wise classification layer. The proposed segmentation network is trained and tested on publicly available dermatology images obtained from challenge host by International Skin Imaging Collaboration (ISIC) in the beginning 2016 on “Skin Lesion Analysis towards Melanoma Detection”. This challenge consists of 900 training sample of dermascopic skin images and 379 for evaluation. Experimental results of proposed segmentation network are very encouraging compare to state of the art result which achieves jaccard index value of 0.928.