基于深度改进卷积神经网络的肺结节图像检测
IV
collected database, and the results showed that the method can effectively detect the pulmonary
nodules in the image compared with the original model.
(2) By studying the classification and identification of lung nodule detection in CT images, it
was found that there were higher false positives in the results of detection and recognition. And by
studying the method of model fusion, it was found that it can effectively improve false positives.
Therefore, this paper adopts the classification model of model fusion convolutional network
structure. After training the model one time, a boosting method was introduced to redistribute the
weights of the misclassified samples, and then the model was trained again until the final test
results are obtained. The experimental results show that compared with the general convolutional
neural network, the proposed model fusion convolutional neural network can effectively reduce
the false positive number of the detection results, and the model achieves higher detection
performance of lung nodules.
(3)Due to the particularity and scarcity of medical images that can be used in practical
research, there is little data available for conducting research and experiments. Through the
research on the generation-adveratial network, it is found that it can effectively copy real samples,
thus effectively expanding the number of research objects. Therefore, this paper analyzes the use
of the method of generation-adveratial network model for data enhancement to improve the
performance of lung nodule detection models for CT images. The model can cut the lung nodule
cubes according to the annotation information, which were sent to the network model to generate
more "true nodules". After the number of positive samples is expanded, they are applied to the
convolution detection model to detect lung nodules. Experiments were carried out using the image
data of the LIDC database as training samples. From the results of the test, the detection rate of the
pulmonary nodules can meet the actual requirements.
Xiaoting Wang (Control Science and Engineering)
Supervised by Professor Tao Gong
KEY WORDS: medical image processing; convolutional neural network; lung nodule
detection; generating confrontation network
万方数据
