Deep Parametric Active Contour Model for Neurofibromatosis Segmentation

Abstract Locating the neurofibromatosis boundaries is one of the most challenging tasks due to the properties of low contrast and irregular blurred boundaries. To handle this problem, we propose Deep Parametric Active Contours (DPAC), a novel end-to-end framework that integrates the precise boundary localization abilities of the Active Contour Models (ACMs) with the robust nonlinear priori feature extraction abilities of the Convolutional Neural Networks (CNN). DPAC employs CNN to predict the parameter maps of the ACM per instance and incorporates them into the energy functional of the ACM to locate the neurofibromatosis boundaries, which makes DPAC end-to-end trainable. We evaluate the proposed method on three neurofibromatosis datasets that we created, and experimental results show that our DPAC method can improve the boundary similarity and precision of the classic ACMs. Even in the case with a small amount of data, our method can achieve good performance outperforming similar methods.