A novel few-shot segmentation framework for computed tomography images

Recently, segmentation based on deep learning has become popular in the medical image domain. However, these deep-learning-based segmentation methods rely heavily on an extremely large number of manual labels and lack the requisite interpretation in mathematical theory. In this paper, we propose a few-shot segmentation framework coupling a deep neural network with active contour models for lung nodules. The advantages of the proposed framework are summarized into 3 categories: 1) The proposed framework requires only few-shot manually labeled samples during network training, which reduces the reliance on data samples and increases the theoretical interpretation with a mathematical model. 2) The initial contours provided by a coarse segmentation network are refined through an optimized active contour model, which offers an automatic process for traditional active contour models without manual setting of initial contours. 3) The proposed framework is easy to extend to other tasks, and the deep learning components are substitutable for plug-and-play presegmented models. In the experiments, the proposed method is evaluated on clinical CT images with lung nodules from a hospital. The results show that the proposed method is effective and robust for improving segmentation accuracy.