Computer-aided pulmonary nodule automatic segmentation method based on neural network

Abstract

The invention discloses a computer-aided pulmonary nodule automatic segmentation method based on a neural network, and belongs to the field of pulmonary nodule automatic segmentation. The method comprises the following steps: preparing original CT image data, and calibrating a lesion area of pulmonary nodule data for training; preprocessing the original CT image data, and intercepting a spatial region of the pulmonary nodule; constructing and training a segmentation model and an objective function of an auto-encoder / decoder and a spatial pyramid pooling structure based on a three-dimensional residual network; and using the trained model to segment the pulmonary nodule of the detection task, and outputting a result of predicting the pulmonary nodule lesion area. By means of the method, a better pulmonary nodule segmentation effect can be achieved, and the method is suitable for pulmonary nodule segmentation tasks in actual clinic.

Description

technical field [0001] The invention relates to the field of automatic segmentation of pulmonary nodules, in particular to a computer-aided automatic segmentation method of pulmonary nodules based on a neural network. Background technique [0002] Lung cancer is a malignant tumor with the highest cancer mortality rate in the world. The five-year survival rate of patients with stage IA lung cancer can reach more than 90%, but once the lung cancer metastasizes, the five-year survival rate of stage IV lung cancer drops rapidly to 5%. Low-dose spiral computed tomography (CT) can screen for curable early lung cancer, so CT has become the most important means of lung cancer screening. As more and more people undergo CT examinations, the burden on doctors to read images is also increasing; secondly, according to the characteristics of CT images, doctors can only associate the information of continuous-level images through their own subjective impressions and experiences. In this w...

AI Technical Summary

Problems solved by technology

[0007] In fact, the pulmonary nodule lesions themselves and their margins, density, texture and other characteristics, as well as the anatomical environment around the pulmonary nodules, have great changes, and some tissues (vessels, solid tissues other than the lung parenchyma) also have similar characteristics. For pulmonary nodules with vascular bundles and pleural traction, traditional machine learning-based pulmonary nodule segmentation methods are difficult to accurately segment such pulmonary nodule regions

Therefore, traditional machine learning methods are difficult to achieve satisfactory results in complex practical application scenarios

[0008] At present, the neural network-based lung nodule segmentation methods are based on the two-dimensional convolutional neural network to segment the lung nodule area slice by slice, even if two-dimensional images with multiple orthogonal perspectives are extracted from the three-dimensional data to extract spatial features. In fact, the rich spatial feature information contained in the CT image is also lost; secondly, it is generally believed that the deeper the network depth, the more feature information can be extracted, while many lung nodule segmentation methods only use convolution- The pooling structure is used to extract the features of pulmonary nodules. Simple stacking of networks in this way can easily cause the problem of gradient disappearance of the neural network during training, thereby reducing the prediction results of the network. In addition, many current methods do not take into account The impact of the noise data introduced by the huge difference in nodule diameter on the feature extraction of pulmonary nodule segmentation makes the segmentation effect also have great differences in the performance of nodules of different sizes; finally, the current work uses the As a pixel-level classification task, the cross-entropy function is used as the objective function. However, due to the data imbalance between the segmented target area and other background areas, it is easy to make the network overfit during the training process.

Images