Invented by Junfeng Wang, Peng Tang, Fan Li, Yihua DU, Yulin JI, Zongan LIANG, Sichuan University
The Sichuan University invention works as followsThe following steps are included in the method: Create the network of fully convolutional networks for lung contour segmentation. Conduct training off-line on the weighting parameter of the network. Read DR images and weighting parameters. Input DR images into the network of fully convolutional networks. Output segmentation results through the network terminal using layer-by-layer feedback. “Establish lung contour according to segmentation results.
Background for Kinds of DR radiography lung shape extraction method based upon fully convolutional networks
The radiography of the chest is the key diagnostic technology for pulmonary diseases. The X-ray is the primary measure for the screening of lung diseases, such as pulmonary inflammation, a lump, tuberculosis and lung cancer. With the development of digital technology, digital radiography, i.e. DR (Digital Radiography), a new technique of digital X ray photography, is gradually replacing the traditional chest perspective method. DR is a technology that uses amorphous flat-panel detectors to directly convert X ray information from the human body to digital signals. The computer then reconstructs the image and performs a series image post-processing. The images are still excellent even if the exposure condition is slightly poor. DR imaging offers high sharpness, low radiation and has become the standard technical device for many hospitals and grassroots physical examination centers.
The principle of DR imaging is that human tissue has different density and thickness. The absorption of X rays varies when the X ray penetrates the different tissues in the body. The X-ray images are formed by the different amounts of X rays that reach the screen. It is used to determine the density and thickness of the tissue, and to speculate and evaluate the diseased area. The thoracic and enterocoelia are the main organs in the human body. They include visceral organs that have high and low densities. The DR radiography exams of a certain size are performed in primary hospitals and medical exam points. However, the DR radiography examinations on a large scale can be very difficult to launch with the existing labor and technological resources.
The basic distance, area and density can be measured by the device or the measurement of the basic distance, area and density; the difficult realization of atopic functions aiming at specific organs or lesions type is due to the fact that the image treatment technology aimed at the tasks and objects still has a lot of technical difficulty in the application layer. It is a measurement of basic distance, area, and density. The application layer of image treatment technology is still very difficult to achieve the atopic functions aiming for specific organs and lesions types.
Contents of the Invention
This invention provides a method of DR radiography lung shape extraction based on a fully convolutional neural network that improves screening treatment efficiency for pulmonary diseases, improves detection accuracy of nidus, and monitors serious infectious disease.
The technical program used in this invention is a method of DR lung contour extraction based on a fully convolutional net, and it comprises the following steps:
Establish a fully convolutional structure for lung contour segmentation
Conduct off-line training for the weighting parameters in the fully convolutional networks;
Read the DR Image and the Weighting Parameters of the Fully Convolutional Network;
Input the DR images into the fully-convolutional network. Output the segmentation result of the image through the terminal network with layer-by-layer feedback. Establish the lung contour according to the segmentation results.
The layer of the network is the unit for the fully convolutional networks. According to the order from input to output, the layers include the data layer.
Further the offline training mentioned on the weighting parameter of the fully convolutional networks comprises the following steps:
Further the following treatment should be performed on the image one time before putting the DR image into the fully convolutional networks:
Convert the DR Image into the floating-point Type Matrix;
Conduct the standardization treatment of the floating-point matrix
Conduct a whitening treatment on the DR image.
The soft maximum value algorithm Softmax was adopted to calculate the overall difference value.
The batch random gradient descent algorithm (Batch-SGD) is used in the calculation to update the parameters of all network layers.
The conversion method for the floating-point matrix mentioned above is to divide the depth of an image (12-bits or 14 bits) in DICOM format, by 212 or214 and convert it into the floating-point matrix.Click here to view the patent on Google Patents.