A POCT-based Image Recognition Study Of Cervical Cancer Cells

Medical resources in remote areas are scarce and many early lesions are difficult to detect.Cervical cancer is one of the most common gynecological malignancies and the HPV vaccine is a good means of prevention,however,the HPV vaccine is not widely available in our country and is expensive.At this stage,cervical cancer screening relies mainly on manual reading,but the complex morphology of cervical cells has resulted in a scarcity of professionals,which is particularly prone to misdiagnosis of false positives and false negatives.A high level of professionalism and intensity of work is demanded of slices reviewers.The number and size of cervical cells can be used as an adjunct to the detection of cervical cancer.Flow cytometry is a standard cellular test,but flow cytometry is expensive and bulky and not suitable for instant detection.In response to the above problems,a POCT-based image detection system for cervical cancer cells is proposed,the system mainly consists of a detection device and related detection algorithms,in which the detection device includes an optical path module,an excitation optical module,a microfluidic chip module and an image acquisition module.A novel watershed image segmentation algorithm and Lt Net image classification recognition deep learning framework were developed using the detection device as a platform.The main elements of the study are as follows.Firstly,a POCT-based cervical cancer cell detection device was designed.Solid Works is used to build 3D models of the optical path module,the excitation module,the microfluidic chip module and the image acquisition module of the detection device.According to the design size requirements,the overall framework of POCT is made by 3D printing technology,the microfluidic chip module is made by microfluidic technology,and the excitation light module is made by laser cutting machine to establish a multimodal instant detection system for cervical cancer cells,which can be identified by the microfluidic chip to count cervical cell solution and cervical cell section for further morphological classification.Secondly,a new method for accurate counting of cervical cancer cells was proposed to identify cervical cancer cells by counting.First,take the PS spheres of different sizes into 95%alcohol to make PS spheres solution,pass into the microfluidic chip,using the designed POCT device on the microfluidic PS spheres solution for image sampling.Second,the acquired images were subjected to grayscale transformation,binarization processing and edge detection.The entropy algorithm,OTSU threshold segmentation and adaptive operator were applied to compare the cell images for binary processing,and Canny operator,Laplace operator,Gaussian Laplace operator and Gaussian differential operator were applied to compare the cell images for edge extraction,and OTSU threshold segmentation and Gaussian Laplace operator edge extraction were selected.Finally,a new watershed algorithm based on OTSU threshold segmentation,Gaussian Laplace edge extraction operator combined with connected region method was proposed,and the improved algorithm achieves 97.3% counting accuracy for PS pellets with higher computational efficiency,which meets the design requirements.Thirdly,a new deep learning classification and identification network was proposed based on morphological recognition of cervical cancer cells.The device was used to collect data from cervical cancer cell sections made by TCT in local hospitals,and combined with public data sets,the images are pre-processed with interpolation,mirroring and RGB conversion,denoising,size scaling,random cropping,etc.Relying on the VOC2012 data set format,the images are annotated manually using Label Me,and a new classification network,Lt Net,is proposed based on the deep network model features of Alex Net,VGGNet,Goog Le Net and Res Net,which can automatically achieve cervical cell classification.Lastly,an experimental platform was set up to test and validate a real-time detection system for multimodal cervical cancer cells.Laboratory-cultured cervical cells were stained according to the FITC fluorescence effect,and microfluidic cervical fluorescent cells were sampled using this device.First,PS glomeruli and cervical cell solution images of different scales were sampled separately from the device using a laboratory inverted microscope system,which confirmed by comparison that the device could meet the test accuracy requirements.Secondly,the improved watershed segmentation method was used to validate the experimental counting of cervical cells at different scales,the test results were compared with flow cytometry with an average accuracy rate of 98.3%,which confirmed that the improved watershed segmentation algorithm was effective for cell counting.Finally,the device was used to image the hospital-treated cervical cell sections,and the improved watershed segmentation algorithm was used to assist segmentation combined with manual screening to obtain cervical cell classification,which was mixed with public datasets for pre-processing operations,and a new designed deep learning framework,Lt Net,was used to perform deep learning training on the classified cells,which was verified by five-fold cross-validation,and finally the TOP-5 average error rate of cervical cells was only 3.35%,which is better than the known classical deep learning classification network framework and meets the experimental requirements and facilitates the automatic screening of cervical cells by machines.