2021/07/26 by Hasib Zunair, Zunair, Hasib, A. Ben Hamza +1 · 2 citations
Computer Science · Medicine · #AI in cancer detection #Radiomics and Machine Learning in Medical Imaging #COVID-19 diagnosis using AI
paper · pdf · doi:10.48550/arxiv.2107.12461
The U-Net architecture, built upon the fully convolutional network, has\nproven to be effective in biomedical image segmentation. However, U-Net applies\nskip connections to merge semantically different low- and high-level\nconvolutional features, resulting in not only blurred feature maps, but also\nover- and under-segmented target regions. To address these limitations, we\npropose a simple, yet effective end-to-end depthwise encoder-decoder fully\nconvolutional network architecture, called Sharp U-Net, for binary and\nmulti-class biomedical image segmentation. The key rationale of Sharp U-Net is\nthat instead of applying a plain skip connection, a depthwise convolution of\nthe encoder feature map with a sharpening kernel filter is employed prior to\nmerging the encoder and decoder features, thereby producing a sharpened\nintermediate feature map of the same size as the encoder map. Using this\nsharpening filter layer, we are able to not only fuse semantically less\ndissimilar features, but also to smooth out artifacts throughout the network\nlayers during the early stages of training. Our extensive experiments on six\ndatasets show that the proposed Sharp U-Net model consistently outperforms or\nmatches the recent state-of-the-art baselines in both binary and multi-class\nsegmentation tasks, while adding no extra learnable parameters. Furthermore,\nSharp U-Net outperforms baselines that have more than three times the number of\nlearnable parameters.\n