Research And Implementation Of Urine Cell Image Segmentation And Recognition Algorithm

The analysis of urine formed cells has important reference value for urine detection,which is one of the three major routine inspection items in medicine.The analysis methods mainly include manual microscopy by phase contrast microscope,and automatic analysis by various types of analyzers.However,there are problems of low efficiency and large error respectively,and it is difficult to deal with the massive formed cell images appearing in urine detection.At present,with the rapid development of deep convolutional neural networks,more and more people are applying these techniques to the analysis of urine formed cells,providing new methods for improving the accuracy of their analysis results.The main contents of the analysis and identification of urine formed cells include 12 types of pathological components in the urine image such as red blood cells,white blood cells,crystals,casts,etc.The process is divided into two important steps,segmentation and recognition.First,the various types of cells in the urine image are segmented and extracted,and then the extracted results are identified and classified.Because the segmentation algorithm and recognition algorithm are less studied in the specific application about urine image,this article will focus on the work of these two aspects based on the characteristics of urine image and cell morphology.Details as follows:(1)On the urine image segmentation,a urine image segmentation algorithm based on edge detection and adaptive seed region growth is designed for the characteristics of urine cells with complex and changeable shapes,large differences in gray values,and inconsistent gradients between cells and background.The algorithm gives the seed selection scheme and growth judgment conditions and growth process when growing the seed region,which can get effective segment for the formed cell of urine image through positioning the seeds on the background of the urine image and selecting adaptive threshold based on the edge detection method.Experimental results show that compared with other segmentation algorithms,this algorithm has better segmentation effect and higher robustness.(2)On the urine formed image recognition,in view of the characteristics of high similarity of individual cells,combined with the feature extraction capability of Goog Le Net and the feature channel learning capability of SENet,a compression algorithm for urine cell image classification algorithm SE-Goog Le Net was proposed.This algorithm learns the dependency relationship between cell feature channels through squeeze operation and excitation operation,and uses feature re-calibration to increase the weight of important features in the current task,which can obtain better classification results.Aiming at the problem of imbalanced cell number distribution,an improved label shuffling method was used for data preprocessing,and a loss function Gauss Loss,which is more suitable for the recognition of urine cell images,was proposed.The loss function can make the model focus more attention on the medium to high hard samples,reduce the impact of mislabeled samples,ensure that the weight of easy samples is not too low,and participate in the gradient optimization of loss functions.Comparative experiments show that the algorithm can effectively improve the accuracy of recognition and classification while guaranteeing the running speed.Based on the proposed urine image segmentation algorithm and the formed cell recognition algorithm,the urine formed cells analysis software is designed and developed.By designing,implementing,and testing the functions,the software can complete the segmentation task of the urine image and the recognition task of segmented formed cell image,and then display the result,which have certain flexibility and good application value.