基于融合先验知识的肺结节深度学习分类方法

Deep Learning Method for Classification of Pulmonary Nodules Based on Fusion of Prior Knowledge

目的提出一种基于融合先验知识的肺结节深度学习分类方法。方法整体模型中包括图像特征提取子模型、语义提取子模型、语义整合子模型、多模态融合部分等。首先通过本文提出算法将医师标注语义信息转换为模糊one-hot码,然后将区域生长法设定不同阈值的输出图像输入语义提取子模型,模糊one-hot码作为多标签训练模型。最后将已训练语义提取子模型固定权重作为语义提取器置入整体模型中,输入图像分别经过图像特征提取子模型和语义提取子模型与语义整合子模型后通过融合输出预测结果。结果以公开数据集LIDC-IDRI作为实验数据做五折交叉验证,得出模型分类性能准确度88.32%、灵敏度81.86%、特异性93.37%、AUC 0.9220。结论基于融合先验知识的肺结节深度学习模型可较高性能实现肺结节良恶性诊断,可作为辅助影像医师诊断的有效工具。

Objective To propose a deep-learning method for classification of pulmonary nodules based on fusion of prior knowledge.Methods The overall model included image feature extraction sub-model(IE model),semantic extraction sub-model(SE model),semantic integration sub-model(SI model),and the part of multimodal fusion.Firstly,the semantic information marked by physicians was converted to fuzzy one-hot codes by the proposed algorithm.Then,the output images of the region-growing method with different thresholds were input into the SE model,and the fuzzy one-hot codes were used as the multi-label to train the model.Finally,the fixed weight of the trained SE model was put into the overall model as a semantic extractor.The input images passed through the IE model,the SE model and the SI model respectively,and the overall model outputted the prediction through fusion.Results The proposed model was evaluated on the public data set LIDC-IDRI by 5 cross-validation,and the experimental results showed that the model archived the performance of accuracy of 88.32%,sensitivity of 81.86%,specificity of 93.37%and AUC of 0.9220.Conclusion The deep learning model based on fusion of prior knowledge for classification of pulmonary nodules can realize the diagnosis of benign and malignant pulmonary nodules with high performance,and can be used as an effective tool for assisting imaging physicians in diagnosis.