Research On Medical Image Segmentation Model Based On Level Set Method And Deep Learning

Medical image segmentation is the premise of medical image understanding and recognition.The quality of segmentation will directly affect the doctor's judgment of disease.However,there are many kinds of medical images,and different medical images have different characteristics,which brings great challenges to image segmentation model.In general,medical images will appear fuzzy edges and uneven intensity.In order to solve these problems,this paper proposes three different medical image segmentation models for different types of medical images.In ultrasonic pupil image,the pupil usually shows different edge intensity in ultrasound image,which leads to edge blur in ultrasonic pupil image,and brings great challenge to image segmentation.In order to solve these problems,this paper assumes that the shape of the pupil is circular,and the pupil segmentation is transformed into finding three points on the pupil.The circle is the contour of the pupil.We use the height of the pupil and the edge below the pupil to narrow the scope of the pupil,which is very helpful for the subsequent selection of pupil points.After reducing the scope of the pupil,this model proposes two operators,convolute the two operators with the pupil region after edge detection,pick out the points whose convolution result is 5 and then formulate corresponding rules to find three points on the pupil.The experimental results show that this model can effectively segment the ultrasonic pupil image.Compared with the traditional active contour model qualitatively and quantitatively,the effectiveness of this model in segmentation of ultrasonic pupil image is proved.Region scalable fitting energy model(RSF)can segment medical images with uneven intensity,but it is sensitive to initial contour and level set parameters.In order to solve this problem,a medical image segmentation model based on level set method and deep learning is proposed in this paper.The model includes length term,data fitting term and restriction term.The constraint term is defined as the error between the level set function and the convolution network(U-Net)used for biomedical image segmentation so that in the segmentation process,the curve will only evolve near the segmentation result obtained by U-Net and the restriction term effectively limits it the evolution of the curve.After that,the split Bregman method is used to quickly minimize the energy functional and the model is used to segment melanoma images.The experimental results show that the model can effectively segment the tumor image,and the segmentation results of the model are better than the traditional active contour model and U-Net.In addition,the segmentation speed of the model is fast,and it is not sensitive to the level set parameters and the initial contour.When the traditional active contour model is used to segment brain tumors,it is easy to segment the brain contour rather than the lesion contour.Therefore,the segmentation results obtained by threshold are used as the initial contour of the RSF model,and the parameters are kept unchanged in the segmentation of brain lesion images by adjusting the parameters.In order to select the threshold automatically,a method of selecting threshold is proposed after observing the histogram of gray image.Then,the segmentation result obtained by this threshold is used as the initial contour,and the gradient descent method is used to minimize the energy functional.The experimental results show that our model can effectively segment brain lesions,and the segmentation results of this model are better than those of threshold segmentation and traditional active contour model.In addition,different threshold experiments show that the threshold automatically selected by the model is not necessarily the best,but it has a good segmentation of brain tumor images.