Deep Learning Method for Classification of Pulmonary Nodules Based on
Fusion of Prior Knowledge
GAO Feng1,2,ZHANG Shirui1
1. School of Precision Instruments and Opto-Electronics Engineering, Tianjin University, Tianjin 300072, China丨
2. Tianjin Key Laboratory of Biomedical Detecting Techniques and Instruments, Tianjin 300072, China
Abstract: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.
高峰1,2,张仕瑞1. 基于融合先验知识的肺结节深度学习分类方法[J]. 中国医疗设备, 2021, 36(3): 54-57.
GAO Feng1,2,ZHANG Shirui1. Deep Learning Method for Classification of Pulmonary Nodules Based on
Fusion of Prior Knowledge. China Medical Devices, 2021, 36(3): 54-57.