| Cancer has always been the biggest medical problem,and compared with the cure rate of60%-70% in developed countries,the survival rate of cancer in China is only 20%-40%.In terms of medical level,China's medical technology don't much different from that of developed countries,However developed countries compared with China have higher early cancer screening rate,and most of them are in the early stage of cancer,which greatly reduces the difficulty of treatment,improves the cure rate and survival time of cancer.In the all cancers,breast cancer is the most dangerous to women,but the cure rate of early breast cancer is up to 95%.Common detection methods of cancer include: blood test,cervical smear,biopsy,gastroenteroscopy,X-ray,B-type ultrasound,molybdenum target photography and etc,most detection methods have specific cancer detection scope.Molybdenum target,B-type ultrasound and biopsy are the most commonly used methods to detect breast cancer in the all cancer detection methods.However molybdenum target photography has some disadvantages,such as radiation,difficult applicability in different scenes,limited detection times and so on.And biopsies can cause physiological trauma to patients,increase the pressure of recovery of patients' focus and the risk of wound infection.Compared with the other two methods,B-type ultrasound has its own advantages in these disadvantages,and it can flexible shooting of images from all angles and parts,so B-type ultrasound is one of the most commonly used methods for breast cancer examination.In conclusion,how to improve the analysis of breast lesions through breast ultrasound image,and improve the diagnosis rate of doctors,as well as lay the foundation for the future machine diagnosis of breast benign and malignant tumors is very meaningful research.In recent years,deep learning has been widely used in medical research,such as arthritis diagnosis,cataract diagnosis,breast cancer-related lymph node metastasis prediction and etc.Artificial intelligence medical image segmentation is also a very hot research topic in recent years.At the same time,many segmentation models based on convolution neural network and full convolution neural network have been proposed,among them have three outstanding models: U-net,CDeep3 M and FFNs will be used as the research of this experiment.The training of models have higher requirements for computing resources,and the training environment of models have their own configuration requirements,so it is very important to train the three models in an independent environment and improve the use efficiency of existing hardware computing resources.In this study,the docker container technology is introduced to integrate each model and the dependency library required by its own environment into a separate container,and any container change will not affect the host and other containers.At the same time,due to the characteristics of docker,it also provides convenience for the research and improvement based on this experiment docker,and provide the portability of the models.The proper segmentation of breast lesion based on mammography is crucial no matter1 for traditional computer aided diagnosis but also for supervised learning by deep learning,which is significant for further quantitative analysis.To achieve a comprehensive diagnosis with versatile deep learning algorithms,a novel Docker based framework is proposed anddesigned by incorporating three popular CNN based models,U-net,CDeep3 M and FFNs.The validation is preceded under a unique amount of B-type ultrasound images.The results indicate that U-net and CDeep3 M have a good capability for breast tumor prediction,while FFNs is incapable for this dataset.Besides,the analysis shows that Docker framework is flexible and extensive for environmental construction for breast cancer real-time prediction,as well as other related applications.The amount of B-type ultrasound dataset and Docker images are released in: https://github.com/Tuer-wsd/Ultrasonography-of-breast. |
PDF Full Text Request