基于深度学习的候选结节检测算法

Nodule candidate detection algorithm based on deep learning

为提高候选结节检测性能,应用深度学习技术提出基于3DSCANet的候选结节检测算法。该算法提出增强坐标注意力机制模块(SCA),在坐标注意力机制的基础上做出改进,使之能提取三维(3D)特征,并引入自适应卷积提取跨通道特征,增加SCA注意力机制的特征提取能力;提出一种将3D长方体锚框转换为3D球体的方法,并进一步引入新的球体交并比损失函数SIoUX,以充分利用肺结节的球体形态特征。在实验阶段,该方法在LUNA16数据集上采用十折交叉验证的方法进行测试,平均召回率CPM达到0.94。

A nodule candidate detection algorithm based on 3DSCANet utilizing deep learning techniques is proposed to improve nodule candidate detection performance.The algorithm employs a strengthen coordinate attention(SCA)module which improves upon the basic coordinate attention mechanism to enable it to extract three-dimensional(3D)features,and incorporates adaptive convolution to extract cross-channel features,thereby enhancing the feature extraction capability of the SCA mechanism.Additionally,a method to convert 3D rectangular anchor boxes into 3D spheres is proposed,along with the introduction of a sphere based intersection over union loss function(SIoUX)to fully leverage the morphological characteristics of lung nodules which are spherical in shape.During the experimental phase,the method is tested on the LUNA16 dataset using ten-fold cross-validation,and it achieves an average recall rate of 0.94.