2018/06/21 by Zahra Mirikharaji, Mirikharaji, Zahra, Ghassan Hamarneh +1 · 1 citation
Medicine · Computer Science · #Cutaneous Melanoma Detection and Management #AI in cancer detection
paper · pdf · doi:10.48550/arxiv.1806.08437
Semantic segmentation is an important preliminary step towards automatic\nmedical image interpretation. Recently deep convolutional neural networks have\nbecome the first choice for the task of pixel-wise class prediction. While\nincorporating prior knowledge about the structure of target objects has proven\neffective in traditional energy-based segmentation approaches, there has not\nbeen a clear way for encoding prior knowledge into deep learning frameworks. In\nthis work, we propose a new loss term that encodes the star shape prior into\nthe loss function of an end-to-end trainable fully convolutional network (FCN)\nframework. We penalize non-star shape segments in FCN prediction maps to\nguarantee a global structure in segmentation results. Our experiments\ndemonstrate the advantage of regularizing FCN parameters by the star shape\nprior and our results on the ISBI 2017 skin segmentation challenge data set\nachieve the first rank in the segmentation task among 21 participating teams.\n