Graph Flow: Cross-layer Graph Flow Distillation for Dual-Efficient Medical Image Segmentation
share
With the development of deep convolutional neural networks, medical image segmentation has achieved a series of breakthroughs in recent years. However, the higher-performance convolutional neural networks always mean numerous parameters and expensive computation costs, which will hinder the applications in clinical scenarios. Meanwhile, the scarceness of large-scale annotated medical image datasets further impedes the application of high-performance networks. To tackle these problems, we propose Graph Flow, a novel comprehensive knowledge distillation method, to exploit the cross-layer graph flow knowledge for both network-efficient and annotation-efficient medical image segmentation. Specifically, our Graph Flow Distillation constructs a variation graph which is employed to measure the flow of channel-wise salience features between different layers. Next, the knowledge included in the variation graph is transferred from a well-trained cumbersome teacher network to a non-trained compact student network. In addition, an unsupervised Paraphraser Module is designed to refine the knowledge of the teacher network, which is also beneficial for the stabilization of training procedure. Furthermore, we build a unified distillation framework by integrating the adversarial distillation and the vanilla logits distillation, which can further promote the final performance respectively. As a result, extensive experiments conducted on Gastric Cancer Segmentation Dataset and Synapse Multi-organ Segmentation Dataset demonstrate the prominent ability of our method which achieves state-of-the-art performance on these different-modality and multi-category medical image datasets. Moreover, we demonstrate the effectiveness of our Graph Flow through a new semi-supervised paradigm for dual-efficient medical image segmentation.
READ FULL TEXT
