基于深度学习的肺部医疗图像识别

Lung medical image recognition based on deep learning

肺结节作为肺癌早期诊断的重要特征,对其识别和类型判断具有重要意义。目前使用迁移学习的识别算法存在着源数据集与目标数据集差距过大问题,对于肺结节特征提取不足,导致效果不佳。故此提出了基于卷积神经网络的改进神经网络模型。将预训练的GooLeNet Inception V3网络与设计的特征融合层结合,提高网络对特征的提取能力;为确定最佳组合方式,对各组以准确率为标准进行测试。实验在LUNA16肺结节数据集上进行。进行分组测试结果表明,改进的网络准确率达88.80%,敏感度达87.15%。在识别准确率和敏感性指标上,与GooLeNet Inception V3算法相比,分别提高了2.72,2.19个百分点。在不同数据集比例下进行实验,同样达到了更优的效果,具有更好的泛化能力。可以给临床诊断提供相对客观的指标依据。

As an important feature of early diagnosis of lung cancer,pulmonary nodules are of great significance to its recognition and classification.At present,there is a large gap between the source data set and the target data set in the recognition algorithm of transfer learning,and the feature extraction of pulmonary nodules is insufficient,resulting in poor effect.This paper proposes an improved neural network model based on convolutional neural network.The pre-trained GooLeNet Inception V3 network is combined with the designed feature fusion layer to improve the network's feature extraction capability.In order to determine the best combination method,each group is tested based on accuracy.The experiment was performed on the LUNA16 dataset of pulmonary nodules.After grouping test,the accuracy and sensitivity of the improved network are 88.80% and 87.15% respectively.Compared with GooLeNet Inception V3 algorithm,the recognition accuracy and sensitivity are improved by 2.72 and 2.19 percentage points respectively.Experiments under different proportions of data sets also achieve better results and have better generalization ability.It can provide relatively objective index basis for clinical diagnosis.