Training FCNs model with lesion-size-unified dermoscopy images for lesion segmentation

Dermoscopy is an important noninvasive diagnostic technique for early detection of melanoma that is the deadliest form of skin cancer. Lesion segmentation is a critical step in analyzing dermoscopy images. An obvious characteristic of the lesions in dermoscopy images is appearing in multi-size, and it becomes a serious barrier to improve the lesion segmentation while using fully convolutional neural networks (FCNs) model, which is a popular tool to carry off this challenging task. In this paper, we propose a method to unify the lesion sizes among various dermoscopy images and use the unified images to train the FCNs model. The method was tested on PH2 and ISIC2017 dataset, and the results of experiments show that using the lesion-size-unified dermoscopy images can greatly improve the segmentation performance. This work is of value on assisting artificial delineation, revealing the relations between performance and various lesion sizes, and guiding significance for preprocessing of input dermoscopy images while using FCNs model.