Automatic liver tumor segmentation plays a key role in radiation therapy of hepatocellular carcinoma. In this paper, we propose a novel densely connected U-Net model with criss-cross attention (CC-DenseUNet) to segment liver tumors in computed tomography (CT) images. The dense interconnections in CC-DenseUNet ensure the maximum information flow between encoder layers when extracting intra-slice features of liver tumors. Moreover, the criss-cross attention is used in CC-DenseUNet to efficiently capture only the necessary and meaningful non-local contextual information of CT images containing liver tumors. We evaluated the proposed CC-DenseUNet on the LiTS dataset and the 3DIRCADb dataset. Experimental results show that the proposed method reaches the state-of-the-art performance for liver tumor segmentation. We further experimentally demonstrate the robustness of the proposed method on a clinical dataset comprising 20 CT volumes.
Lane detection is a critical technology in autonomous driving, which requires both high detection accuracy and real-time efficiency, especially in complex scenarios like severe occlusions, and dazzling light conditions. To achieve the above requirements, based on re-parameterization and attention mechanism we propose a real-time lane detection network with a classification-based manner named RepATT. A new backbone Extended RepVGG (ExtRep) which subtly decouples the training time and inference time is proposed as the feature extractor. Specifically, the ExtRep has two additional branches of varying complexities and scales compared to the original RepVGG. This enhancement improves the performance without any additional inference cost by enhancing the representational capacity of single convolution with re-parameterization methods. In addition, by integrating the Convolutional Block Attention Module(CBAM), our model improves the performance in the no-visual evidence scenarios. CBAM generates the attention maps via channel and spatial dimensions enabling the model to concentrate more on the lanes. Through comprehensive experiments conducted on CULane and TuSimple datasets, the results demonstrate our proposed method achieves the fastest inference speed, which surpasses the baseline model by 24% to 363 FPS with lower computing cost, while the F1-measure of RepATT also outperforms the previous methods on both datasets.