Research On Multi-Disease Tumor Region Segmentation Based On Convolutional Neural Network

Cancer has brought great pain and harm to normal people.The current cancer lesions have become an important factor in the death of residents in our country and even the world.Observing the Ki-67 slide with a microscope,calculating its positive rate,and analyzing its cancerous symptoms is a crucial step.Due to the limited number of doctors’ readings in a short period of time and the diversity of the shape of the cancer lesion area,the location is not fixed,and each doctor’s personal experience will produce a different diagnosis,which makes it difficult to diagnose cancer lesions.Taking digital pathological images as an example,based on deep learning for image segmentation,researchers use the source data to make a suitable data set,and use the data set they made for training,and continuously adjust some impact factors in the training code,Make the test effect better,train a more accurate model for tumor region segmentation.In this way,when the model is put into use,it can not only improve the speed of doctor diagnosis,but also improve the level of intelligent medical treatment.In view of the above problems and background,we use breast cancer,colorectal cancer,and neuroendocrine tumor data sets of three disease types to study the segmentation of tumor regions.The thesis completed the following work:1.Using the cGAN network,the representative sub-function in pix2pix-semantic segmentation,segment the tumor regions of the three diseases.On the one hand,based on the similarity of human tissue structure and shape,the data of the three diseases are mixed and trained.On the other hand,the data of the three diseases are trained separately.Both are trained based on the conditional generation of the adversarial network.The experimental results show that the segmentation accuracy dice value fluctuates around 0.7.Compared with the data of poorly differentiated tumor regions,the well-differentiated tumor region segmentation effect is better,indicating that the pix2pix network multi-disease hybrid training has great application value to the tumor region segmentation.2.Using the LinkNet network with encoder-decoder structure,combined with the model-based transfer learning method,the LinkNet urban landscape segmentation model is used as a pre-training model to segment the tumor areas of the three diseases.The training is conducted in two modes.On the one hand,the data of the three diseases are mixed for training,and on the other hand,the data of the three diseases are trained separately.The experimental results show that the dice values of the segmentation accuracy of multi-discipline hybrid training and single-discipline training both oscillate around 0.8.Compared with the pix2pix network results,the segmentation effect of LinkNet network is relatively good.