Research On Cell Segmentation Of IMC Images With Deep Learning

Imaging Mass Cytometry(IMC)is a highly multiplexed antibody imaging technology that can simultaneously image multiple(up to 40)proteins on tissue slices at single-cell resolution,providing more possibilities for comprehensive exploration of the tumor microenvironment.Therefore,accurate segmentation of single cells has become a prerequisite for effective analysis of the tumor microenvironment.However,due to existing multiple channels of markers in IMC images,how to process such high-dimensional image data has become the biggest problem in IMC image segmentation tasks.Many methods of segmentation for IMC images have been developed,but they generally do not work well or require a lot of time to manually label them.Therefore,this paper proposes a set of segmentation methods based on deep learning for the segmentation of IMC image data.Compared with the existing methods,the algorithm proposed in this paper achieves more accurate and efficient segmentation results.The main research work is as follows:(1)Based on the convolutional neural network and existing segmentation methods with advantage on dividing clumped cells,a new method is proposed for cell segmentation of highly multiplexed tissue images.First,the pixel classification probability map generated by interactive machine learning is used to guide the training of multiple deep neural networks through the method of knowledge distillation,and the model with the best performance is selected as the general pixel classifier.The pixel classifier classifies each pixel in the cell image into nucleus,cytoplasm,and background.Finally,use CellProfiler to post-process the classification probability map to achieve cell segmentation result.(2)In addition,to make our model applicable to all IMC images,we combine certain position-specific proteins into two position information channels,which represent the nucleus and cytoplasm/membrane respectively.(3)Since the purpose of cell segmentation is cell quantification,the key point to verify the effectiveness of this method is whether it can produce more accurate cell phenotype measurements.Therefore,in addition to many accuracy indicators,this thesis also proposes a set of validation methods for cell segmentation.By measuring the cell phenotype in images,we analyze and verify the reliability of segmentation results from multiple angles such as cell morphology,area overlap,proteins expression.