Unsupervised Optic Disc Segmentation for Cross Domain Fundus Image Based on Structure Consistency Constraint

摘要

Deep convolution neural networks (DCNNs) are playing critical roles in various computer vision tasks, though they also suffer from many problems. The performance of DCNNs will severely degrade when facing gaps between the training set. and test set. It is quite common in the field of medical image analysis where domain shift between images acquired from different devices or patients. A feasible solution to this problem is collecting more labeled data from the testing set for fine-tuning DCNNs to improve the generalization ability. Unfortunately, medical image labeling is complicated and time-consuming, which makes fine-tuning hard to apply. In this paper, we propose a novel network model which makes full use of the existing pre-trained network models, effectively reducing domain shift and restraining performance degradation. This method dramatically reduces the dependence on labeled data and achieve better results even if trained with a small amount of unlabeled data. We transform the testing images (target domain) with adversarial learning mechanism and make them look similar to training data (source domain) that can be segmented directly by the pre-training model trained on the training set. In addition, we introduce additional structural consistency constraints to suppress the distortion during forwarding propagation when the model is trained with fewer samples, which ensure the structure of the generated image is consistent with the input image. We validate our method in retinal fundus image segmentation task. Experiments show that the proposed method suppresses the degradation of model performance caused by domain migration, and achieves almost the same segmentation performance as the original training data.