Research On Cancer Cell Segmentation Based On CNN

Cancer cells are a kind of abnormal cells,and it is the source of cancer.Segmentation of cancer cells is an important and effective method to quantify the malignancy of tumor.It is extremely important for understanding the mechanism of causing disease,increasing the efficiency of drug development,improving the diagnosis and treatment of diseases.Cell segmentation in microscopic images based on deep learning(DL)technology has received strong attention in the field of biomedical engineering,and has become one of the focus studied and front field.Considering the difficulties in segmentation of cancer cells,such as small,high density,adhesion occlusion,different shapes and sizes,and low signal-to-noise ratio of microscopic images,some segmentation methods are proposed for cancer cells in microscopic images based on CNN in this thesis,and it mainly includes the following research work:(1)For some problems,such as small target size and different shape of cancer cells,a multi-scale convolutional block(MC-Block)is proposed for cancer cell segmentation in microscopic images by adopting multiple different scale convolutions.The proposed MC-Block is applied for testing in various network architectures,such as U-Net,VGG19 and Res Unet.And the results show that the introduction of MC-Block can significantly improve network performance.(2)To deal with the high density of cancer cells,a density loss function is constructed by using cell density information as a supervised signal.The proposed method can effectively improve the segmentation precision of cancer cell under the condition of the occlusion and high density by integrating the density estimation algorithm into the MC-Unet.(3)For the cell adhesion,a boundary attention segmentation network is proposed by selecting the boundary contour as a supervisory signal,and it is integrated into the MC-Unet architecture by constructing a boundary branch loss function.The results shows the proposed method can effectively strengthen cell boundary features,thereby improving the performance of cell segmentation.To verify the effectiveness of the above methods,six pathological cell image datasets with different characteristics are selected for testing.The results show that the proposed methods can effectively solve the problems of small,high-density cancer cells,adhesion occlusion,different shapes and sizes,and low signal-to-noise ratio in microscopic images.