Abstract: As a novel machine learning method, Relevance Vector Machine（RVM） has drawn many scholars’ attention. The recently developed Multi-Kernel Learning RVM（MKLRVM） method makes it a more preferable classifier in the field of pattern recognition. The setup of the kernel parameters and the combination weights between kernel functions in MKLRVM plays an important role in classification accuracy, while they are usually empirically determined rather than a quantitative method. To address this issue, the Particle Swarm Optimization（PSO） and the Second-Order Cone Programming（SOCP） are introduced to find the optimal parameters, reasonable combination of kernel functions is constructed, and a quick calculation method is given. Finally the improved MKLRVM is applied in lung nodule detection using the CT images from the publicly available LIDC database. Through the image processing module, the features of nodule candidates are extracted. The proposed MKLRVM is employed to classify the nodules, and experimental results demonstrate that the computation efficiency is improved and better classification performances are achieved by the presented PSO and SOCP based MKLRVM method, compared to that of the PSO based MKLRVM.

Key words: lung nodule detection, Relevance Vector Machine（RVM）, Particle Swarm Optimization（PSO）, Second-Order Cone Programming（SOCP）

摘要： 在模式识别问题中，相关向量机（RVM）作为一种新的机器学习方法备受关注，近年来，多核RVM方法的提出使得RVM得到更广泛的应用。多核RVM模型中核参数的取值及不同核函数组合权重系数的取值对模型分类性能至关重要，然而在实际应用中其值却多由经验值给定而非定量分析计算得到。为此，对基于粒子群算法（PSO）及基于二阶锥规划（SOCP）的多核RVM参数优化模型进行研究，构造合理的核函数组合，并给出快速求解方法。最后将该方法应用到肺结节检测中，采用公共数据集LIDC中的肺部CT图像，通过图像处理模块，提取候选结节的特征信息，利用改进的多核RVM模型对肺结节进行分类验证。实验结果表明，与基于PSO的多核RVM模型相比，基于PSO与SOCP相结合的多核RVM模型不仅提高了运算效率而且取得了更好的分类性能。

关键词: 肺结节检测, 相关向量机, 粒子群优化, 二阶锥规划