EdgeMedNet: Lightweight and Accurate U-Net for Implementing Efficient Medical Image Segmentation on Edge Devices

Convolutional neural networks have gained tremendous success in computer vision and medical imaging applications. To make these models truly portable and compatible for prototyping, their efficient implementations on low-power devices are vitally imperative. Thus, in this brief, we propose EdgeMedNet, which is one lightweight and accurate U-Net model to enable the efficient medical image segmentation on Intel/Movidius Neural Compute Stick 2 (NCS-2). Firstly, we optimize the feature maps of the baseline U-Net model into the optimum point for striking a balance between parameters reductions and segmentation performance. Then, we propose the enhanced activation block to provide comprehensive multi-scale analysis, for facilitating avoiding the representational bottleneck. By utilizing these elementary units, we further introduce the cascaded attention module and customize the layers to construct our EdgeMedNet. Extensive experiments are conducted on three medical imaging datasets, including BraTs dataset of brain MRI, heart MRI dataset, and COVID-19 CT lung and infection segmentation dataset. The experimental results demonstrate that we can gain up to
$6.1\times $
reductions of the model parameters. Compared with the state-of-the-art works, our EdgeMedNet is capable of achieving as high as
$2.0\times $
inference speedup on NCS-2, while maintaining the segmentation performance.