【摘要】 结肠镜检查广泛应用于结直肠癌的早期筛查和诊疗，但仅靠人工判读结肠息肉漏检率较高，有研究统计可达25%。基于深度学习的计算机辅助技术有助于提高息肉检测率，但目前深度学习的主流分割网络U-Net存在着两个局限：一是编解码的输出特征图之间存在着语义鸿沟；二是U-Net的双层卷积单元无法学习多尺度信息；割裂地看待容易使模型陷入局部最优。提出一种基于同构化改进的U-Net网络，不仅能缓解编解码特征间的语义鸿沟，且具备提取多尺度特征的能力。首先，在U-Net编解码器和跳层路径中，引入同构单元IU构成同构网络I-Net,以减少编解码器之间的语义鸿沟；接着，结合密集连接和残差连接的优点，设计密集残差单元DRU以学习多尺度信息；最后，将同构网络的处理单元初始化为密集单元，构成基于密集残差单元的同构网络DRI-Net。使用包含612幅结直肠镜息肉图像的公开数据集CVC-ClinicDB,采用5折交叉验证评估所提出的模型，DRI-Net可得Dice系数为90.06%,交并比(IoU)为85.52%,与U-Net相比，Dice系数提升8.50%,IoU提升11.03%。此外，在国际ISIC2017皮肤镜挑战数据集上验证模型在其他模态数据的泛化性，2 000幅训练，600幅测试，获得的Dice系数为86.57%,IoU为79.20%,与ISIC 2017排行榜第一名的方法相比，Dice系数提升1.67%,IoU提升2.70%。实验表明，DRI-Net能有效解决U-Net存在的局限，且泛化性良好。更多还原

【Abstract】 Colonoscopy is a widely used technique for colon screening and polyp lesions diagnosis. Nevertheless, manual screening using colonoscopy suffers from a miss rate around 25% of polyps. Deep learning-based computer-aided diagnosis(CAD) for polyp detection has potentials of reducing the human errors. Polyp detection depends on encoder-decoder network(U-Net) for polyp segmentation. However, U-Net has two limitations, one is that the semantic gap exists between the feature maps from the encoder and decoder; the other one is convolutional layers in the encoder-decoder processing units fail to extract multi-scale information. In this work, we proposed an identical network(I-Net) to tackle the problems in a consolidated manner. The I-Net introduced identical units(IU) both in skip connections and encoder-decoder sub-networks of U-Net to reduce the semantic gap. Meanwhile, motivated by the dense and residual connections, we designed a dense residual unit(DRU) to learn multi-scale information. Finally, DRI-Net was developed by initializing IU to DRU, which not only alleviated the semantic gap between the encoder and the decoder but also learned multi-scale features. We evaluated the proposed methods on the CVC-ClinicDB dataset containing 612 colonoscopy images through five-fold cross validation. Experimental results demonstrated that the DRI-Net achieved Dice coefficient of 90.06% and intersection over union(IoU) of 85.52%. Compared to the U-Net, DRI-Net improved the Dice coefficient of 8.50% and IoU of 11.03%. In addition, we studied the generalization of the proposed methods on International Skin Imaging Collaboration(ISIC) 2017 dataset including a training set of 2 000 dermoscopy images for model training and a test set of 600 images for model evaluation. The study indicated that the I-Net achieved Dice coefficient of 86.57% and IoU of 79.20%. Compared to the first-place solution on ISIC 2017 leaderboard, the DRI-Net improved Dice coefficient of 1.67% and IoU of 2.70%. In conclusion, the results demonstrated that DRI-Net effectively overcome the limitations of U-Net and improved the segmentation accuracy in the polyp segmentation task, and showed the great generalization capability on other modality data.更多还原