Liver segmentation in abdominal CT images via auto-context neural network and self-supervised contour attention

Objective: Accurate image segmentation of the liver is a challenging problem owing to its large shape variability and unclear boundaries. Although applications fully convolutional neural networks (CNNs) have shown groundbreaking results, limited studies focused on performance generalization. In this study, we introduce CNN for abdominal computed tomography (CT) images that focus generalization accuracy. Methods: To improve performance, initially propose an auto-context algorithm in single CNN. The proposed network exploits effective high-level residual estimation obtain prior. Identical dual paths are effectively trained represent mutual complementary features accurate posterior analysis liver. Further, extend our by employing self-supervised contour scheme. We sparse penalizing ground-truth more attentions failures. Results: used 180 CT training validation. Two-fold cross-validation presented comparison with state-of-the-art networks. experimental results show better accuracy when compared reducing 10.31% Hausdorff distance. Novel multiple N-fold cross-validations conducted best network. Conclusion significance: method minimized error between test than any other modern Moreover, scheme was successfully employed introducing self-supervising metric.