The Research And Application Of Convolution Neural Network In Pathological Diagnosis

Pathological diagnosis plays a crucial role in the diagnosis of cancer.I In the pathological diagnosis,a small amount of tissue is cut from the patient’s tumor to make a pathological section.The pathologist observes the section under the microscope and gives an assessment report of the benign and malignant tumors and the degree of metastasis based on the tumor cells and their tissue morphology.Call it a biopsy(abbreviation: biopsy).Pathological diagnosis is an important basis for benign tumors and cancer diagnosis.The accuracy and accuracy of the results are particularly important for patients and will directly affect the next treatment plan.Therefore,pathological diagnosis is also called the "gold standard" in disease diagnosis,and its authority is far superior to other imaging examinations such as CT images,nuclear magnetic images,etc.Because the accuracy of pathological diagnosis depends heavily on the diagnostic level of pathologists,the results from different time pathologists are different.Therefore,combined with computer-aided diagnosis system,the accuracy of pathological diagnosis will be effectively improved.With the rise of computer,Internet and large data,deep learning has made many breakthroughs in the fields of image recognition,autopilot and speech recognition.In this paper,we use deep learning technology combined with pathological images to study the algorithm of deep learning in pathological diagnosis,and take the breast cancer pathological diagnosis as an example,design the corresponding diagnosis algorithm and diagnosis system.The main research content is as follows:1)Full-convolutional neural network(FCN)is used to study the segmentation algorithm of deep learning on pathological images.Combined with the multi-resolution features of pathological section microscopy,a multi-resolution,multi-network probabilistic fusion diagnostic algorithm was proposed to achieve the diagnosis of pathological images.Through experiments and comparison with other diagnostic algorithms,the algorithm of this paper has high accuracy and robustness.2)A set of training methods using deep learning for the diagnosis of pathological images was studied.Compared with natural images,pathological sections have ultra-high resolution and are pyramid-shaped.The underlying high-resolution images can be observed in cell structure.The upper low-resolution images reflect the organizational morphology.When using a full convolutional network for training The image at different resolutions is also completely different,and the network convergence is slower and it is not even possible to train normally.In this paper,the methods of migration learning,proportion extraction,and re-fitting of poorly-fitted data sets are used in this paper to improve the convergence speed of the network and also fully learn the training set.3)Using Ubuntu operating system and GPU graphics acceleration,a set of pathological image diagnosis system for breast cancer was designed.The pathological images of breast cancer were divided into four and two types of segmentation were performed.The diagnostic accuracy was high and the algorithm was validated.