Orchestrating Medical Image Compression and Remote Segmentation Networks

Deep learning-based medical image segmentation on the cloud offers superb performance by harnessing the recent model innovation and hardware advancement. However, one major factor that limits its overall service speed is the long data transmission latency, which could far exceed the segmentation computation time. Existing image compression techniques are unable to achieve an efficient compression to dramatically reduce the data offloading overhead, while maintaining a high segmentation accuracy. The underlying reason is that they are all developed upon human visual system, whose image perception pattern could be fundamentally different from that of deep learning-based image segmentation. Motivated by this observation, in this paper, we propose a generative segmentation architecture consisting of a compression network, a segmentation network and a discriminator network. Our design orchestrates and coordinates segmentation and compression for simultaneous improvements of segmentation accuracy and compression efficiency, through a dedicated GAN architecture with novel loss functions. Experimental results on 2D and 3D medical images demonstrate that our design can reduce the bandwidth requirement by 2 orders-of-magnitude comparing with that of uncompressed images, and increase the accuracy of remote segmentation remarkably over the state-of-the-art solutions, truly accelerating the cloud-based medical imaging service.