Research On Medical Image Segmentation Method Based On Machine Learning

Computer-aided diagnosis of medical images can provide efficient and accurate analysis results for doctors,and become an important tool for clinical pathological diagnosis.With the widely use of machine learning methods in computer vision,many research results have been made in the field of medical image processing.Cervical cell image segmentation is important for cervical cancer screening.This paper did deep research on cervical cell images segmentation.At the same time,the method used in cervical cell image is also applied to other medical images to prove the effectiveness of machine learning for medical image segmentation.First,this paper uses the U-net network learning method to segment the cell image,and divide the cervical cell image into three parts:background,cell and nucleus.In order to obtain more feature information,this paper improves the U-net network structure,expands the receptive field,and makes the segmentation result more accurate.Then,use watershed and area analysis method to find independent cells in the image.Then,by comparing various methods,this paper chooses multiple active contour methods to segment overlapping cells.The active contour method initializes a contour for each cell through the distance between the pixel and the cell nucleus,and combines the prior information of the shape of the cell,the edge information of the image and the mutual information between different contours to establish the energy functional of the level set function.Minimize the energy functional to obtain the cell outline and complete the segmentation of each cell.To sum up,this paper combines deep learning in machine learning and traditional watershed and active contour methods to complete cell segmentation in cervical cell images,and can complete accurate extraction of every cells including independent cells and overlapping cells.The Precision coefficient is 0.9317,and the DICE coefficient is 0.8820.In addition,the proposed improved U-net network is used for fundus images,and the segmentation indicating that the method in this paper can be applied to medical image segmentation and has great auxiliary value for pathological screening.