用于肺结节分类的3D稠密倒残差潜在均衡注意力网络

3D Dense Inverted Residuals Latent Equilibrium Attention Network for Pulmonary Nodules Classification

摘要: 目前基于深度学习的高性能的肺结节良恶性分类网络通常由于结构复杂，伴随着大量的参数量和计算量需求。为此本文提出了3D稠密倒残差潜在均衡注意力网络(3D Dense Inverted Residuals Latent Equilibrium Attention Network，简称3D DIRLEAN)，以实现在进一步提高肺结节分类精度的同时减少模型的参数量节省实际应用内存。所提出的3D DIRLEAN网络框架仅需4个阶段，主要由3D稠密倒残差LEA模块(3D DIRL)和潜在均衡注意力模块(LEA)组成。我们受生物神经元的响应机制启发构建了LEA。它通过我们设计的潜在均衡能量函数直接估计三维权重，获取更具区分性的特征；运算过程0参数并且可以有效减少网络层之间的响应滞后，加速模型收敛。此外，我们构建的3D DIRL结合了倒残差和稠密连接的优势，将其同时用于特征处理。它使用稠密连接获取不同网络层的特征信息，结合倒残差实现特征重用，最大限度利用特征信息加强网络的判断能力；同时，标准卷积的分解为网络降低了参数量和计算量。这两个模块共同作用使3D DIRLEAN能够以最少的参数量和计算量实现最高的分类精度。所提出的方法应用于LUNA16数据集。实验结果表明，3D DIRLEAN的分类精度达到了93.33%，并且参数量和计算量分别降低到了2.85M、0.63G，总体上优于先进的同类方法。

Abstract: Presently, the development of advanced classification networks for accurate differentiation between benign and malignant pulmonary nodules, utilizing deep learning techniques, often entails a multitude of parameters and substantial computational demands. Consequently, this research proposes the innovative 3D Dense Inverted Residuals Latent Equilibrium Attention Network (3D DIRLEAN) to improve the accuracy of pulmonary nodule classification and to optimize the number of parameters in the model, thereby saving the memory of practical applications. 3D DIRLEAN mainly includes the 3D Dense Inverted Residual LEA Module (3D DIRL) and the Latent Equilibrium Attention (LEA) mechanism. The LEA module directly estimates three-dimensional weights through a bespoke latent equilibrium energy function, thereby capturing more distinct features. The operation process employs zero parameters, effectively curtailing response lags between network layers and hastening model convergence. Furthermore, the constructed 3D DIRL component seamlessly integrates the advantages of inverted residuals and dense connections for concurrent feature processing. Capitalizing on dense connections, it acquires feature information across diverse network layers, complemented by inverted residuals to enable feature recycling, thereby optimizing the utilization of feature data. Simultaneously, the deconstruction of standard convolutions serves to curtail the model’s parameter count and computational overhead. The experimental accuracy of 3D DIRLEAN on LUNA16 reached 93.33%, and the parameters and FLOPs were reduced to 2.85M and 0.63G respectively. Clearly, LEA and 3D DIRL work together to enable the 3D DIRLEAN framework to achieve the highest classification accuracy over similar advanced technologies, while minimizing parameter counts and floating-point operations (FLOPs).