Automatic Segmentation Of Female Pelvic Endangered Organs Based On Deep Learning Fusion Network

Cervical cancer is one of the common gynecological malignancies.The incidence rate is second only to breast cancer,which has become a major hidden danger to women’s health.Radiation therapy,as the main treatment for cervical cancer,is of great significance for patients with cervical cancer,especially advanced patients.With the development of radiotherapy equipment,Intensity Modulated Radiotherapy(IMRT),Tomotherapy and Volume Modulated Radiation Therapy(VMAT)have been applied to the treatment of cervical cancer.One of the requirements of these techniques to ensure the accuracy and accuracy of treatment implementation is to accurately delineate organs on the patient’s planned CT image.However,the sketching relies mainly on the experience and abilities of the doctor.There are often differences in the organs at risk that are drawn by different doctors.The results of the same doctor’s sketching at different times may be slightly different.And the process of delineation takes a long time,which will prolong the waiting time of patients.Therefore,radiotherapy clinically urgently needs an efficient and accurate automatic delineation of software that threatens organs.At present,the Atlas-based Segmentation(ABS)method is commonly used in radiotherapy.This method first establishes a map library by using CT data of a moderately shaped patient and other patient’s delineation data.When a new patient is to be drawn,the delineation result can be obtained by registering the contour of the most matching CT image in the library to the new CT.However,this method is limited by the patient’s body shape,library size,CT image layer,registration accuracy,etc.,and the clinical effect is not good.In particular,the pelvic organs,their size,texture grayscale and relative position are greatly different among different individuals due to natural variability,disease state,treatment interference and organ deformation,resulting in extremely low clinical acceptability.With the success of convolutional neural networks in the field of image understanding and analysis,they have been applied to the processing of radiotherapy medical images.However,the design and training of the automatic segmentation model of pelvic crisis organs have the following three difficulties: 1.The size and shape,texture gray scale and relative position of organs such as the intestines and bladder are subject to individual differences,filling states,disease states and The interference of treatment leads to poor consistency of training data,and it is difficult to increase the depth of neural network in order to avoid the occurrence of underfitting phenomenon during training.2.The adhesion of pelvic organs to peritoneum and mesentery is low,and it is difficult to distinguish,so it is difficult to distinguish.Manual labeling is difficult,resulting in low data quality;3.Clinical women with pelvic tumors have less CT data,making feature extraction and deepening the depth of the network more difficult.Therefore,a network model for solving the problems of degradation,gradient disappearance,etc.of 3D convolutional neural network optimization when small training samples are needed is needed.At present,the realization of automatic debridement of pelvic endangered organs,especially the good delineation effect of small intestine organs,is still a research hotspot and difficulty in the field of medical image processing based on deep learning.Taking advantage of the two network structure models of Dense Net and V-Net,this paper proposes a novel Dense V-Network algorithm based on multi-model fusion,which tries to maximize the likelihood of training data when the sample size is small.Quickly and accurately delineate the pelvic endangered organs and verify them.The experimental results show that the Dense V-Network network model can accurately delineate the five dangerous organs of the bladder,small intestine,rectum,femoral head and horsetail when the training sample is small.The quantitative evaluation by DSC value is better than the single model,which proves the advantages of multi-model fusion.However,if applied to radiotherapy clinical work,it still needs doctor supervision.With the optimization of the algorithm structure and the increase of the training data,it is expected that the sketching result can be further improved.