基于深度学习的结直肠腺体自动分割算法研究

AUTOMATIC COLORECTAL GLAND SEGMENTATION ALGORITHM BASED ON DEEP LEARNING

为实现腺体自动化分割,减轻病理学医生的工作量,帮助医生做出更加准确的临床决策,提出一种基于注意力机制和可变形卷积的适合腺体分割的深度神经网络模型(Adaptive-Gland-Segmentation-Net,AGS-Net)。该模型使用分组卷积和注意力机制使模型具有更强的表征能力,增加可变形卷积层以适应不同分化程度的腺体形状。在GlaS数据集上,加入染色标准化预处理的AGS-Net在检测结果、分割性能和形状相似性等三方面与竞争方法相比,具有很大的优势。

In order to realize automatic gland segmentation,reduce the workload of pathologists and help doctors make more accurate clinical decisions,an adaptive-gland-segmentation-net(AGS-net)based on attention mechanism and deformable convolution is proposed.In this model,grouping convolution and attention mechanism were used to make the model more representative.A deformable convolution layer was added to adapt to the glands with different levels of differentiation.In GlaS dataset,the performance of AGS-Net with stain normalization ranked in the top three of the existing algorithms in terms of detection results,segmentation performance and shape similarity,and it had great advantages.