An Enhanced Deep Learning Framework for Skin Lesions Segmentation

Reliable and accurate segmentation of skin lesions images is an essential step in analysing skin lesions for the clinical diagnosis and treatment of melanoma skin cancer. Skin cancer analysis and detection has been automated over the years using various computing techniques and algorithms. Machine learning techniques such as deep learning methods have also been recently applied in diagnosing the disease. Segmentation identifies the shape of the features and the region of interest for analysis. Inconsistency in the delicate arrangement of skin lesions, coupled with possible presence of noise and artefacts such as hairs, air or oil bubbles on skin lesions, weak edges, irregular and fuzzy borders, marks, dark corners, skin lines and blood vessels on skin lesions has made automation of skin lesions segmentation challenging. The proposed deep learning framework is composed of a deep convolutional neural network with an encoder-decoder type architecture that fully integrates a dice coefficient loss function and employs elastic transformation techniques for data augmentation. The multi-stage segmentation approach adopted in this work learns contextual information by extracting discriminative features at the encoder stage of the system and also captures the object boundaries of the skin lesions images at the decoder stage. This enable the system to effectively segment the challenging and inconsistent skin lesion images. This system is further improved with the combination of effective data augmentation technique and the dice loss function. The performance evaluation of the proposed model with evaluation metrics such as Dice Coefficient, Jaccard index, Accuracy and Sensitivity gives improved and promising results when compared with some existing state-of-the-arts techniques.