Research On Cell Detection Algorithm Based On Key Point

Deep learning technologies have been widely employed in the medicine field and has made significant progress in clinical applications.Being the basic component and structural unit of organism,cells can reflect the physiological state of tissues and organs.Quantitative analysis of cells and cell structure based on image analysis techniques has important clinical value for pathological diagnosis,cancer identification,disease classifi-cation and prognosis.At present,there are the following challenges in deep learning-based cell detection research:（1）Compared with other medical images,cell images usually contains a large number of cells with blurred contours in overlapping and clustered distributions.The complexities of the cell image,i.e.,diverse imaging models,staining techniques,and data heterogeneity also bring challenges for developing deep learning-based models for cell im-age analysis.（2）The existing researches on cell detection focus on predicting the location of cells by center points,while the size,aspect ratio and morphology are all significant for cell image analysis.（3）The lack of labeled data is a common problem in develop-ing medical image analysis models,because the conflict between the difficulty in medical data labeling and the data starvation of deep learning hinders the development of efficient,accurate and practical deep learning schemes for cell image analysis.To address the above issues,this research explores the feasibility and effectiveness of the key point based cell image analysis model.The main research work are as follows:This research propose a cell detection algorithm based on keypoints（KPCDetector）,which utilizes keypoint annotations to construct an object detection model for predicting the location information of cell objects.First,an hourglass network-based keypoint de-tection network is established to predict the cell center point and edge points probability heatmap.Then,the peak extraction algorithm is employed to predict the coordinates of center and edge points.We propose to utilize the geometric position relationship to ob-tain the cell detection boxes for quantifying the location information and morphological features of cells.Furthermore,this study proposes to construct the cell graph according to the coordi-nate position of edge points,and designs the detection performance optimization module GMM for morphological constraints according to the cell matrix to optimize the perfor-mance of cell detector using the cell morphological feature constraint algorithm.This study validates the effectiveness of the algorithm in multiple cellular object detection sce-narios,including,fluorescence images,histopathological images,and single-cell parasite images.In addition,this thesis also demonstrates the process and mechanism of convo-lutional neural network in cell feature extraction through visualization technology,and proves the importance of morphological features in convolutional neural network.Fur-thermore,this study constructs the cell graph model according to the coordinate position of edge points,and calculates the cell matrix and cell geometric description,and uses the cell geometric feature constraint algorithm to optimize the performance of cell detector.Finally,this study uses visualization technology Grad-CAM and Guided Grad-CAM vi-sualizes the mechanism of convolution neural network and discusses the importance of geometry in this process.This research conducts experiments in four cell detection scenarios to evaluate the performance of the proposed algorithm.KPCDetector obtains the detection results of77.96 m AP in Kaggle2018 fluorescent cell image dataset,51.37 m AP in electron micro-scope image dataset,and 48.88 m AP in histopathological image dataset.Additionally,KPCDetector achieves 37.14,32.70,and 58.72 m AP in the dataset of three types of uni-cellular parasites.