基于Involution算子和协调反向注意力的息肉图像分割

Polyp Segmentation Based on Involution and Coordinate Reverse Attention

结肠影像中的息肉存在形态多变、边缘粘膜模糊的特点。针对现有的结肠息肉分割网络因卷积算子的固有限制而导致的特征提取不充分,以及因区域和边界关系刻画不全导致的分割不完整的问题,提出一种基于Involution算子和协调反向注意力的网络模型IN-CRNet。在编码器部分,设计基于Involution算子的感受野模块InRFB,自适应地捕获不同尺度的上下文语义信息,增强模型对形态复杂的结肠息肉的识别能力;在解码器部分,设计协调反向注意力模块CRA,同时关注区域和边界的重要程度,并构建两者的关系,自底向上地逐渐完善边缘轮廓的分割细节。在5个结肠息肉数据集上的验证实验表明,IN-CRNet具有良好的分割精度和泛化能力。

Polyps in colon images are characterized by variable morphology and blurred edges.Aiming at the problems of the current neural networks for polyp segmentation,such as the inadequate feature extraction due to the inherent limitations of convolution,and the unsatisfactory segmentation due to the incomplete relationship between area and boundary,a network(IN-CRNet)based on Involution and coordinate reverse attention was proposed.In the encoder,an Involution-based Receptive Field Module(InRFB)was designed to adaptively capture contextual information at different scales.It improved the ability to detect complex and variable polyps.In the decoder,a coordinate reverse attention module(CRA)was designed to focus on the importance of both regions and edges and establish the relationship between them.It gradually refined the details of the edges from the bottom to up.The experimental results on five public datasets show that IN-CRNet effectively improves the accuracy of segmentation and has good generalization ability.