A lung lobe segmentation method and system based on three-dimensional convolutional neural network

Abstract

The invention discloses a lung lobe segmentation method and system based on a three-dimensional convolutional neural network. The method includes the following steps: constructing a training image data set for lung lobe segmentation; constructing a lung lobe segmentation network based on a three-dimensional convolutional neural network and performing network training, Preprocess the training image data set, and output the category probability map of each pixel after the training is completed; use the Dice Loss loss function to calculate the loss of the category probability map to which each pixel belongs, and weight the losses of multiple category probability maps to obtain the total loss; Set the weight attenuation and learning rate attenuation, train the network until the network converges; input the image to be tested to the trained lung lobe segmentation network after preprocessing and output the prediction result; restore the prediction result to the original input size of the image to be tested after post-processing to obtain the final Split results. The invention can obtain the lung lobe segmentation result after preprocessing and network reasoning, realizes the end-to-end design, and improves the efficiency and precision of the lung lobe segmentation.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method and system for lung lobe segmentation based on a three-dimensional convolutional neural network. Background technique [0002] At present, lung lobe segmentation methods are roughly divided into four strategies: methods based on image registration, methods based on watershed segmentation, methods based on supervised learning, and methods based on deep learning. The existing methods based on image registration are more dependent on structure Map and computational complexity; the accuracy of existing methods based on watershed segmentation depends on the constructed feature map, which is prone to mis-segmentation problems due to unclear boundaries of watershed; the accuracy of existing methods based on supervised learning depends on artificial The extracted features require a large number of labeled samples; the existing deep learning-based methods first segment t...

AI Technical Summary

Problems solved by technology

[0002] At present, lung lobe segmentation methods are roughly divided into four strategies: methods based on image registration, methods based on watershed segmentation, methods based on supervised learning, and methods based on deep learning. The existing methods based on image registration rely more on structure Atlas and large amount of calculation; the accuracy of the existing methods based on watershed segmentation depends on the constructed feature map, which is prone to mis-segmentation problems due to unclear watershed boundaries; the accuracy of existing methods based on supervised learning depends on artificial The extracted features require a large number of labeled samples; the existing deep learning-based methods first segment the 3D bounding box of the lung area by cascading, and then perform lung lobe segmentation on the 3D bounding box, although the lung lobe segmentation accuracy is improved. , but the end-to-end workflow is not implemented, and the workflow is more complicated

Images