Cell Detection Method For Digital Pathological Images Based On Deep Learning

Pathological examinations by observing the structure of cells and lesions under the microscope is the"gold standard"for the diagnosis of most diseases,and the final diagnosis of many diseases depends on the analysis results of pathological morphology.However,traditional pathology diagnosis and scarce pathologist resources have been unable to meet the increasing demand for pathological diagnosis,and this contradiction has become an important factor in the stagnation of pathology development in China.Therefore,cell detection of digital pathology images through big data and artificial intelligence technology,it can provide aids and references for pathologists’diagnostic analysis and make fast,accurate and non-subjective diagnostic results,and improve the accuracy and efficiency of pathology diagnosis to make up for the shortage of pathologists to alleviate the clinical needs of pathological diagnosis in primary hospitals.With the development of digital pathology,cell detection of digital pathology images using artificial intelligence has become a research hotspot,but the problems of multi-scale cells,imbalance between foreground and background and cell adhesion in digital pathology images have led to the accuracy and detection effect of the model cannot reach the clinical level of application to pathology.To solve the above problems,this paper proposes an adaptive anchor region proposal network based on sample weights and a non-maximal value suppression algorithm based on cell instance density,and the detection effect of the proposed method is verified on a real clinical dataset of a hospital.Finally,the digital pathology image assisted diagnosis system is designed and implemented.The research content of this paper is mainly divided into the following three parts:（1）An adaptive anchor region proposal network based on sample weights is proposedConsidering the problem of multi-scale cell instances and imbalance between foreground and background samples in digital pathology images,the Sample Weight Learning and Adaptive Anchor Region Proposal Network（SWLAA-RPN）model based on sample weights is proposed in this paper.Firstly,a bootstrap anchor module is introduced to adaptively generate non-uniform and arbitrarily shaped anchor points to improve the detection of cell instances at different scales.The sample weight learning module is also added to alleviate the problem of the imbalance between foreground and background samples through sample weighting,so that the model pays more attention to the foreground sample data related to the cell instance,and eliminates the sample data of the tissue fluid background area that is not related to the cell.The experimental results show that the SWLAA-RPN network model effectively solves the detection problem in digital pathology images,and the model achieves better detection effect and accuracy.（2）A non-maximum suppression cell detection method based on cell instance density is proposedConsidering the problem of highly dense cell adhesion scenes in digital pathology images,a non-maximum suppression algorithm based on cell instance density is proposed,which solves the problem of missed detection and false detection in the cell adhesion scene.Firstly,the algorithm calculates the density value idensity of each cell instance,and then obtains a dynamic threshold N_density to suppress the detection frame according to the threshold function.Then the penalty function of Soft-NMS is introduced to suppress the detection score by weighting.The method can dynamically adjust the threshold value according to the density of cell instances.In the dense cell adhesion scene,the threshold will dynamically increase,while in the sparse scene,the threshold will dynamically decrease.The experimental results also show that the method improves the detection performance better than other non-maximal suppression algorithms,and also adapted to other object detection models.（3）Designed and implemented a digital pathology image assisted diagnosis systemBased on the above models and methods,this paper implements a digital pathology image assisted diagnosis system,which is developed by separating the front and back ends.The main purpose of the system is to verify the effectiveness of the cell detection method proposed in this paper and apply the method to actual clinical scenarios to assist doctors in pathology diagnosis.The system mainly includes basic function module,pathologist function module and system administrator function module to help pathologists manage patient’s information and digital pathology images,and use the system to perform manual annotation or automatic annotation tasks on patient’s pathology images.