Recognition Of Acid-fast Bacilli Images In Cerebrospinal Fluid Based On Convolutional Neural Network

Objective As a diagnostic method of tuberculous meningitis,improved acid-fast staining has high specificity,convenient and quick operability,low biosafety requirement,and still plays an important clinical role in the diagnosis of acid-fast bacilli associated infection.However,because acid-fast bacilli are relatively small under the microscope,and the concentration of acid-fast bacilli in the cerebrospinal fluid is usually very low,it is very difficult to find the acid-fast bacilli manually under the microscope,which requires a lot of time for careful search,and requires high technical requirements for the examiners.This study explored the feasibility of using YOLOv3 algorithm to detect acid-fast bacilli images in cerebrospinal fluid slides,in an attempt to improve the detection rate and inspection efficiency of the improved acid-fast staining technique.Methods First of all,the records of patients with positive modified acid-fast staining in our hospital from January 1,2004 to December 31,2017 were statistically analyzed and evaluated,and the acid-fast bacilli images in the positive slides of modified acid-fast staining in cerebrospinal fluid were collected.Secondly,after we decided to use modified acid fast stain alveolar lavage slides instead of cerebrospinal fluid slides,to collect images from a sufficient number of acid fast bacilli as training,to train can identify acid bacillus image convolution neural network algorithm based on YOLOv3,identify and test the algorithm in the test group alveolar lavage slides and cerebrospinal fluid in the glass acid bacillus image precision ratio and recall ratio and F1-score.Results In the first part,the records of positive slides with modified acid-fast staining of cerebrospinal fluid were reviewed and the following deficiencies were found:1.Although positive patients found the number of bacteria that met the positive criteria(acid-fast bacilli>or equal to 3),quite a number of slides were not completely read;2.Second,the record of acid-fast bacilli distribution in and out of CSF cells is unclear.Our research group tried to accumulate enough images of acid-fast bacilli in cerebrospinal fluid as the training group,but the number of acid-fast bacilli in cerebrospinal fluid slides was rare,so it was impossible to accumulate enough pictures of acid-fast bacilli in cerebrospinal fluid for machine learning in a short time.In the second part,after two independent algorithm training,the YOLOv3 algorithm obtained for the recognition of 105 bacterial images in the alveolar lavage fluid has an accuracy rate of 93.0%,a recall rate of 90.4%,and a F1-score of 0.91.The accuracy,recall and F1-score of the proposed algorithm were 88.0%,76.1%and 0.81,respectively,indicating that the YOLOv3 algorithm had good accuracy and robustness in identifying acid-fast bacilli in CSF slides after modified acid-fast staining.Conclusion The YOLOv3 algorithm trained by sufficient number of acid-fast bacilli images can approach the level of artificial recognition of acid-fast bacilli,and the detection efficiency is significantly improved.