Study On The Key Technology Of Automatic Assurance Of Cone-Beam CT Image Quality Based On Deep Learning

Image-guided radiotherapy(IGRT)is currently the most important means of precision radiotherapy.It is used for initial positioning and radiotherapy placement,providing positional information for precise positioning of lesions and tumor irrational,and is widely used in radiotherapy therapy.IGRT relies on the quality of the cone-beam computed tomography(CBCT)system to verify the patient positioning and develop monitoring of radiotherapy therapy,which directly affects the accuracy of the IGRT process.The quality of cone-beam CT image is affected by a variety of factors,and its quality evaluation index refers to the linear CT value,high resolution,low-contrast resolution and uniformity.The key technology for the automatic evaluation of four indicators is the automatic positioning of each inserts in the CTP404 module of Catphan 504 phantom.The premise of automatic positioning is to segment each inserts.The segmentation technology is mainly the traditional methods and machine learning.The CTP404 module has multiple inserts with different densities,and the number of inserts pixels accounts for 8% of the total image in the axis slice image(image size is 512×512).Considering that the segmentation target is multi-class and has certain requirements on the accuracy,deep learning and traditional methods will be combined.The main research content of this paper is as follows:(1)To explore the U-Net model structures and an improved model based on U-Net.Although the U-Net model achieves multi-objective segmentation,the training accuracy in the verification phases of network training and the learning of low-level features need to be further improved.On the basis of U-Net model,the residual block and the recurrent convolutional layer are added to the forward convolutional layer to improve training accuracy and recognition degree.Therefore,U-Net with forward convolutional layers with residual connectivity is used,which is called residual U-Net(ResU-Net),the architecture is U-Net with forward recurrent convolutional layers,which is named RU-Net,the last architecture is U-Net with recurrent convolution layers with residual connectivity,which is named R2U-Net.(2)The four network models were trained and tested by the Catphan 504 phantom collected from the CBCT system,of which 252 images were used for training and 63 images were used for testing.According to the evaluation indexes dice similarity coefficient and F1-Score,the target inserts of Air 、 PMP 、 LDPE 、 Delrin and Teflon are quantitatively analyzed,and the segmentation results of various models are compared.The results of R2U-Net were the best among the four models.Then the edge of the R2U-Net model result is processed by the Moore neighborhood tracking algorithm,and more accurate results are obtained to achieve the positioning of inserts.