Skin lesion segmentation using object scale-oriented fully convolutional neural networks

Melanoma is the deadliest form of skin cancer, and its incidence level is increasing. It is important to obtain a diagnosis at an early stage to increase the patient survival rate. Skin lesion segmentation is a difficult problem in medical image analysis. To address this problem, we propose end-to-end object scale-oriented fully convolutional networks (OSO-FCNs) for skin lesion segmentation. Given a single skin lesion image, the proposed method produces a pixel-level mask for skin lesion areas. We found that the scale of the lesions in the training dataset affects a large number of the segmentation results of the lesions in the testing phase, and thus, a training strategy called object scale-oriented (OSO) training is proposed. First, the pre-trained network of VGG-16 is adapted and is transformed into fully convolutional networks (FCNs). Second, after very simple preprocessing, skin lesion images with boundary-level annotations are fed into the FCNs for fine-tuning training based on the pre-trained model using OSO training. During the OSO training, the training dataset is divided into 2 subsets according to an index called the object occupation ratio, and then the whole training dataset and the 2 subsets are used to train 3 different scale-oriented FCNs. A dataset provided by the International Skin Imaging Collaboration (ISIC), ISIC2016, is used for training and testing. Our algorithm is compared with the state-of-the-art algorithms, and the experimental results demonstrate that the segmentation accuracy of our algorithm is higher or very close to the performances of the other algorithms.