Nuclei Segmentation Of Breast Cancer Pathological Images Based On Deep Learning

Breast cancer is the phenomenon that mammary epithelial cells proliferate out of control under the action of a variety of carcinogenic factors,and its incidence ranks the first among female malignant tumors.The Nottingham grading system,which is the most widely used,evaluates three scores: glandular duct formation,nuclear diversity and mitosis count.Because mitosis are sparse and the background is complex,it is very difficult for segmentation and recognition,so it is the most difficult to count mitosis.The use of automatic method to complete the mitosis counting has a positive significance for assisting doctors in diagnosis and reducing the workload of doctors.However,in the actual environment,various automatic methods based on deep learning require a large amount of data for training and verification,and better hardware resources to cope with the huge computing overhead.Based on the above content,this paper carries out the following research work.First,work with the hospital to produce a new dataset using clinical data.Secondly,in order to obtain high precision of mitosis segmentation and recognition,a method of mitosis segmentation and classification based on lightweight residual and attention gate is proposed.Finally,considering the requirements of practical applications and the characteristics of small target segmentation,further research on model lightweight and receptive field feature fusion direction is carried out,and a multi-branch feature pyramid based mitosis segmentation method LKSIPP＿Net is proposed for mitosis segmentation.The specific work and innovation of this paper are as follows:(1)A breast cancer pathological image mitosis segmentation and detection dataset GZMH is prepared based on clinical image data.The original data are provided by Ganzhou Municipal Hospital of Guangdong Provincial People’s Hospital and carefully marked by professional doctors.The SC＿Mitosis proposed in this paper is used to verify the dataset.(2)In SC＿Mitosis,the depthwise separable convolution is used to construct the residual structure to reduce the number of parameters,and a gated structure is proposed to integrate the channel and spatial attention in the GRU,so as to construct the segmentation network segmentation candidate region.And use Res Net34 classification network and get the final classification results.Finally,the effectiveness of the proposed method is verified on the ICPR 2012 competition dataset and GZMH dataset.Among them,F1-score on ICPR2012 dataset reaches 0.8796,and F1-score on larger scale GZMH dataset reaches 0.5685,both of which are superior to other comparison methods,showing good performance of SC＿Mitosis.(3)In LKSIPP＿Net,a multi-branch feature pyramid using large convolution kernel and internal convolution is proposed to enhance semantic extraction and fusion capabilities,and a segmentation network is constructed by combining residual structure.Adaptive averaging pooling with involution is used to enlarge the receptive field,and parallel depthwise separable structure is used to enlarge the receptive field with large convolution kernel while reducing the number of parameters.The experimental results show that the F1-score of LKSIPP＿Net on the ICPR 2012 dataset reaches 0.8173,which is better than other methods for comparison,showing a good potential for research based on the characteristics of the receptor field.