Mitosis Detection In Breast Cancer Based On Weakly Supervised Learning

The pathological images of breast cancer contain rich phenotypic information,which plays an indispensable role in the diagnosis and treatment of breast cancer.The Nottingham grading system is the basis for the diagnosis of breast cancer,and mitosis count is one of the most commonly used diagnostic methods.In clinical practice,pathologists usually need to manually calculate the number of mitotic cells from a large number of breast cancer pathological images.This process will take a lot of time and the diagnosis is subjective.In addition,pathologists usually use dot labels when marking mitotic cells.This label is a weak label,which is different from the conventional label used in the target detection model.With the help of the objective analysis ability of weakly supervised learning technology in pathological image detection and the ability of processing weak label information,this paper proposes a method of breast cancer cell mitosis detection based on weakly supervised learning for automatic detection of mitosis.The main work of this paper is as follows:1.This paper uses deep convolutional neural networks to realize automatic detection of mitosis when only weak label information is available.In this paper,the central label provided by the ICPR 2014 dataset is extended to an elliptical label to solve the problem that the central label lacks enough information to directly train the detection model.Since the appearance of mitotic cells in breast cancer pathological images is similar to that of other cells,the number of mitotic cells is far less than the number of non-mitotic cells,and the size of mitotic cells is relatively small,this article uses an end-to-end multitasking network framework Mask R-CNN to automatically detect mitosis,the feature pyramid network and residual network 101 in the network can extract multi-scale mitosis features.In this paper,Mask R-CNN is used as the benchmark model,the ellipse label is used to train the benchmark model,and the parameters of the model are set according to the characteristics of breast cancer pathological images.The precision is 0.59,the recall is0.295,and the F-score is 0.394.2.This paper further proposes a breast cancer cell mitosis detection method based on weakly supervised learning to improve the missed detection problem of the benchmark model.This method adds four different attention mechanisms on the basis of the benchmark model,which mainly involve the cascade of channel attention,spatial attention,channel attention and spatial attention.This paper uses ellipse labels to train each improved model to explore the impact of different attention mechanisms on the performance of the detection model.The experimental results show that the cascade of channel attention and spatial attention can improve the sensitivity of the detection model to mitotic features and provide more information for the mitosis detection model.Its precision is 0.625,the recall is 0.568,and the F-score is 0.595.Realize the effective detection of mitosis in the pathological image of breast cancer.In this paper,we use weakly supervised learning technology to automatically detect mitosis,provide help for the diagnosis of breast cancer and reduce the workload of pathologists.This task has practical significance in clinical application of medicine and is helpful to promote the development of anti-cancer process.