SD-UNet: Stripping down U-Net for Segmentation of Biomedical Images on Platforms with Low Computational Budgets

During image segmentation tasks in computer vision, achieving high accuracy performance while requiring fewer computations and faster inference is a big challenge. This is especially important in medical imaging tasks but one metric is usually compromised for the other. To address this problem, this paper presents an extremely fast, small and computationally effective deep neural network called Stripped-Down UNet (SD-UNet), designed for the segmentation of biomedical data on devices with limited computational resources. By making use of depthwise separable convolutions in the entire network, we design a lightweight deep convolutional neural network architecture inspired by the widely adapted U-Net model. In order to recover the expected performance degradation in the process, we introduce a weight standardization algorithm with the group normalization method. We demonstrate that SD-UNet has three major advantages including: (i) smaller model size (23x smaller than U-Net); (ii) 8x fewer parameters; and (iii) faster inference time with a computational complexity lower than 8M floating point operations (FLOPs). Experiments on the benchmark dataset of the Internatioanl Symposium on Biomedical Imaging (ISBI) challenge for segmentation of neuronal structures in electron microscopic (EM) stacks and the Medical Segmentation Decathlon (MSD) challenge brain tumor segmentation (BRATs) dataset show that the proposed model achieves comparable and sometimes better results compared to the current state-of-the-art.