Enhancing U-Net with Spatial-Channel Attention Gate for Abnormal Tissue Segmentation in Medical Imaging

Abstract

In recent years, deep learning has dominated medical image segmentation. Encoder-decoder architectures, such as U-Net, can be used in state-of-the-art models with powerful designs that are achieved by implementing skip connections that propagate local information from an encoder path to a decoder path to retrieve detailed spatial information lost by pooling operations. Despite their effectiveness for segmentation, these naïve skip connections still have some disadvantages. First, multi-scale skip connections tend to use unnecessary information and computational sources, where likable low-level encoder features are repeatedly used at multiple scales. Second, the contextual information of the low-level encoder feature is insufficient, leading to poor performance for pixel-wise recognition when concatenating with the corresponding high-level decoder feature. In this study, we propose a novel spatial-channel attention gate that addresses the limitations of plain skip connections. This can be easily integrated into an encoder-decoder network to effectively improve the performance of the image segmentation task. Comprehensive results reveal that our spatial-channel attention gate remarkably enhances the segmentation capability of the U-Net architecture with a minimal computational overhead added. The experimental results show that our proposed method outperforms the conventional deep networks in term of Dice score, which achieves 71.72%.

Deep learning-based detection and segmentation of diffusion abnormalities in acute ischemic stroke

“…In addition, the t-score method performed worse on our testing dataset (Dice about 0.37) than what is described by the developers 14 (Dice about 0.5) Therefore, the classical t-score method was considered insufficiently efficient to segment lesions in our large and heterogeneous clinical dataset. The architecture of our proposed 3D DAGMNet, depicted in Figure 2, is equipped with intra skip connections as UNet3+ 24 , fused multi-scale contextual information block, deep supervision, L1-regularization on final predicts, Dual Attention Gate (DAG) [25][26][27] , self-normalized activation (SeLU) 42 , and batch normalization. The details of the important components and training techniques/parameters are outlined in the following subsections.…”

“…Further developments of UNet variants, such as Mnet, DenseUnet, Unet++, and Unet3+ [22][23][24] optimized the features utilization. The emergence of attention-gate techniques [25][26][27] conditioned networks to focus on local semantical features. Recent studies applied "attention UNets", for example, to predict final ischemic lesions from baseline MRIs 28,29 .…”

Enterprise imaging and big data: A review from a medical physics perspective

“…However, over the past few years supervised learning for biomedical images segmentation has managed to achieve a human-level performance that is very promising [101]. In particular, the U-Net architecture and training strategy [103] which was originally proposed to deal with the lower number of samples commonly found in biomedical domains has been successfully applied to many segmentation problems, including abnormal tissue segmentation [104], organ segmentation in CT images [105], and tumour segmentation in brain MRI [106].…”

“…The use of attention mechanisms and saliency masks have gained some traction in this area, as they provide a way to visualize what region of an image was attended to that led to the predicted outcome. They were employed recently in models used to screen chest X-rays of COVID-19 patients [119], predict lung module malignancy from longitudinal CT [120], perform abnormal tissue segmentation in natural, CT and MRI images [104] and quantification of knee osteoarthrisis in X-ray images [121], while the MDNet model of Zhang et al [109] used attention mechanisms to indicate which area of the image corresponded to the text in the generated diagnostic report.…”

Monaural Speech Enhancement with Complex Convolutional Block Attention Module and Joint Time Frequency Losses

“…The self-attention focused on the response of positions in a sequence while CCBAM focuses on cross-channel and spatial information of feature maps. Another related work is the spatial-channel gating scheme proposed for the U-Net structure [18], which addressed medical image segmentation problem. [19] also applied a concurrent space-channel-wise attention to the redundant convolutional encoder-decoder (RCED) for speech enhancement.…”