基于U-Net改进的内窥镜息肉图像分割算法

An improved endoscopic polyp image segmentation algorithm based on U-Net

摘要:

息肉图像的分割在临床医疗和计算机辅助诊断技术等领域具有广泛的研究和应用价值，但是就目前的研究和应用需求来看，准确的息肉分割仍然是一项挑战. 针对内窥镜息肉图像中出现的息肉与黏膜边界不清晰、息肉的大小和形状差异较大等影响分割质量的问题，该文提出了一种基于U-Net改进的息肉图像分割算法（SBF-Net）. 首先，在U-Net架构上引入了边界特征加强模块（BFEM），考虑到息肉边界和内部区域的关键线索，该模块利用编码器高层特征生成额外的边界补充信息，在解码器阶段进行融合，提升模型处理边界特征的能力. 其次，该模型的解码器（GFBD）采用了从上至下逐步融合特征的方式，将编码器阶段的输出特征经过局部加强（LE）模块之后再逐步融合边界特征，这种多尺度特征融合方式有效缓解了编码器和解码器之间的语义差距问题. 最后，在后处理阶段采用测试时数据增强（TTA）来进一步对分割结果进行细化. 该模型在CVC-300、CVC-ClinicDB、Kvasir-SEG、CVC-ColonDB和ETIS-LaribPolypDB等5个公开数据集上进行了对比实验和消融实验，实验结果证明了该文所改进方法的有效性，并在内窥镜息肉图像上表现出更好的分割性能和更强的稳定性，为息肉图像的处理和分析提供了新的参考.

Abstract:

The segmentation of polyp images has extensive research and application value in the fields of clinical treatment and computer-aided diagnostic technology, but accurate polyp segmentation is still a challenge in terms of current research and application needs. In order to solve the problems that affect the segmentation quality of endoscopic polyp images, such as the unclear boundary between polyps and mucous membranes, and the large difference in the size and shape of polyps, this paper proposed an improved U-Net polyp segmentation algorithm. Firstly, the boundary feature enhancement module was introduced on the U-Net architecture. Considering the key clues of polyp boundary and internal area, this module used the high-level features of the encoder to generate additional boundary supplementary information, which is fused at the decoder stage to improve the ability of the model to process boundary features. Secondly, the decoder of the model adopts the method of gradually fusing features from the top to the bottom. After the output features of the encoder stage are passed through local emphasis module, the boundary features are gradually fused. This multi-scale feature fusion method effectively reduces the semantic gap between the encoder and the decoder. Finally, test-time augmentation was used in the post-processing stage to further refine the segmentation results. The model has been compared and ablated on five public datasets: CVC-300, CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB and ETIS-LibPolypDB. The experimental results prove the effectiveness of the modified method, and it shows better segmentation performance and stronger stability in the endoscopic polyp image, which provides a new reference for the processing and analysis of the polyp image.