| The lung is the largest organ in the human body,and it is also an organ that is constantly exposed to the external environment.With the improvement of medical infrastructure,statistic data about disease is becoming more and more abundant and accurate.Statistics as of 2018 show that lung cancer has already become the main cause of human death among cancers.Clinical screening and treatment of lung cancer are mainly achieved by doctors observing CT images.However,due to the variations of lesion in CT images,it has always been a challenge for doctors to accurately identify potential lesions by observing CT images.It is also a very time-consuming procedure.Therefore,how to use computers for automated disease detection has become a hot research area.Lesion detection in the early stage usually requires to design a large number of features manually,and feeds the features to a traditional machine learning algorithm for detection.Such a method relies heavily on the effectiveness of hand-crafted features.With the rapid development of deep learning,the use of convolutional neural networks for automatic feature extraction replaces the procedure of manual feature design.The use of neural networks with a large number of parameters to automatically detect potential lesions in CT images has become the main direction of disease detection.The CT images of the lungs are 3D data.In the scanning process of the CT images,the use of different scanning machines and different physical parameters of the scanning machines may lead to large difference in the thickness of the scanned images.Using3 D convolutional networks to detect lesion on the 3D data directly will result in pool performance due to the difference in the thickness.Aiming at the problem that the 2D model cannot use the slice context information and the 3D model performs poorly on data with large slice thickness differences,we proposes a 2.5D lesion detection model for 3D CT images based on the vanilla FCOS.1)The slice grouping design is introduced which inputs multiple adjacent slices into the model in groups,and a 2D convolution network is used to extract the structural features of adjacent slices,which avoids using a 3D backbone network for feature extraction.It reduces the computational complexity and the impact of slice thickness differences on model performance.2)An adaptation module is designed that uses 3D convolutional layers to extract the slice-level features,and then the feature compression along the slice axis is performed to get a 2D feature map which contains 2.5D structural features,so that the 2D task branch structure of FCOS can be directly reused.3)A adaptive weighting loss function is proposed.The regression branch loss is weighted with centerness adaptively.Adaptive weighting loss improves the position accuracy of the samples near central area of the bounding box,and the regression branch loss is smoother.4)In order to make full use of the pre-training model and reduce the model’s resource occupation,a parameter correction processing is also adapted during model initialization.5)Mixed precision training is used to further reduces the memory usage of the model.Based on the above design,we established a 2.5D FCOS lesion detection model.In order to verify the effectiveness of the model,we design detailed experiments from several different perspectives.The experimental results on the tianchi lung CT dataset show that our model is higher than the classic Faster R-CNN model in FROC 10.1%.By increasing the number of input slices,the performance of the model can be improved from 0.388 to 0.419.By changing the number of input slices,the model can balance accuracy and efficiency.Compared with single-precision training,mixed-precision training can reduce the memory usage of model training by about 24%,and reduce the time consumption of a single iteration from 12.4% to 18.4%.From the experimental results,it can be seen that our model has higher performance on the lung 3D CT images,and it also has greater advantages in deployment. |
PDF Full Text Request