| Human body fluids,secretions and excretions such as blood,leucorrhea,urine and feces are routine detection items in physical examination.At present,the manual microscopic examination is the primary method for detecting the formed elements such as cells in most medical institutions,but it has low detection efficiency and the detection results depend on the experience of doctors.With the development of in vitro diagnostic(IVD)equipment and deep learning technology,it is an inevitable trend that the automatic and intelligent sample analyzer gradually replaces manual detection.The autofocus,segmentation and recognition of formed elements in complex biological samples are the important research topic to realize the intellectualization of full-automatic sample analyzer.This dissertation takes the human feces and leucorrhea samples as examples to analyze the theoretical difficulties and key technologies deeply.Based on deep learning technology,this research develops autofocus method and the intelligent segmentation and recognition method of formed elements for complex biomedical microscopic images.The research contents and main achievements of this dissertation are as follows:1.Obtaining clear microscopic images is the essential precondition of segmentation and recognition of formed elements.Aiming at the problem that liquid stratification in human feces and leucorrhea sample solutions make it difficult for the traditional sharpness algorithms to find the clearest image during the focusing process,a convolution neural network(CNN)based on local maximum gradient method and mask attention mechanism is developed in this dissertation,which is called GMANet(gradient mask attention network).The local maximum gradient image is introduced into the convolution layer as the mask image,in order to suppress the characteristics of fuzzy regions or blank backgrounds and make the network pay more attention to the clear regions.The experimental results show that the average accuracy of GMANet in predicting the clearest fecal microscopic image is 97.7%,which is more than 10%higher than the traditional sharpness algorithms.The average detection time of GMANet for detecting one fecal microscopic image is 99 ms,which meets the needs of real-time detection.The GMANet has good generality and the model trained on the focus image dataset of feces can be directly used for autofocus of leucorrhea and blood samples without transfer learning,with an average accuracy of 94.99% and 99.08%,respectively.Therefore,the developed GMANet solves the autofocus problem of complex biomedical microscopic image effectively and can be integrated into full-automatic sample analyzer.2.Aiming at the problem of missed detection of formed elements caused by the liquid stratification in sample solution,inspired by manual microscopic examination,a formed elements detection method based on super depth of field is developed.In this dissertation,the clearest image and its adjacent images obtained in autofocus process are combined along the channel dimension as the input of the CNN,and the super depth of field information can be learned through the morphological changes of formed elements at different positions in axial direction to improve the discrimination ability of the network.In order to improve the recognition effect of the detection network on small objects,SE(Squeeze-and-Excitation),GAD(Global Attention Down-Sample)and DCN(Deformable Conv Nets)are added into the feature extraction stage.The experimental results show that the m AP(mean average precision)value of the developed detection network on test set of fecal microscopic image is 0.881,which is more than 10% higher than current mainstream object detection methods based on deep learning.The average detection time for recognizing the formed elements in one visual field is 98 ms,which meets the requirements of real-time detection.The formed elements detection method based on super depth of field can be integrated into the full-automatic feces sample analyzer.3.Aiming at the problem that the formed elements detection method based on super depth of field has poor detection effect on the test set of leucorrhea microscopic image,this dissertation developes a new formed elements detection method based on object relation module and super depth of field,which is called OR-SDo F-Net(Object Relation and Super Depth of Field Network).Firstly,the clearest image and its adjacent images obtained in autofocus process are combined along the batch dimension as the input of the CNN.After passing through the feature extractor,they are separated along the batch dimension and used to establish feature pyramids,respectively.Each feature pyramid contains the deep semantic features of corresponding microscopic image.The detection network establishes the association between multiple feature pyramids through the improved object relation module and then fuses multiple feature pyramids,so that the detection network can learn the super depth of field information of formed elements through the deep semantic features of microscopic images.In order to improve the discrimination ability,the detection network uses the balanced feature pyramid method to re-establish the connection between different feature layers of fused feature pyramid.The experimental results show that the m AP value of OR-SDo F-Net on test set of leucorrhea microscopic image is 0.863,which is more than 6% higher than current mainstream object detection methods.The average detection time for recognizing the formed elements in one visual field is 177 ms,which meets the requirements of real-time detection.The OR-SDo F-Net can be integrated into the full-automatic leucorrhea sample analyzer.4.The process of TV(Trichomonas vaginalis)staining detection method is complex and some stains can destroy the activity and structure of formed elements.Therefore,the staining detection method is not suitable for the full-automatic leucorrhea sample analyzer.Considering the motion characteristics of active TV,a feasible scheme is to record videos of leucorrhea samples with a microscopic camera and recognize TV according to motion information.However,most TV is far away from the focus plane where the clearest image is captured,causing it to be defocused and blurred in video images,appearing as moving shadow regions.The morphology of blurred TV is similar to the background and the motion characteristics are not obvious,which makes it difficult to identify.Aiming at the problem of blurred TV recognition,this dissertation develops a rough-fine two-step TV region segmentation method and two improved TV segmentation networks: rough segmentation network based on optical flow information and fine segmentation network based on mask recovery.The rough segmentation network uses the image information and the optical flow information between adjacent frames as the input to segment the TV region roughly.The fine segmentation network uses the image information and the rough segmentation result as the input,and outputs the fine segmentation result of TV region after contour correction.The experimental results show that the combination of two segmentation networks can effectively extract the blurred TV regions and realize the TV recognition.The average Io U(Intersection over Union)value of the developed method on the test set is 72.09%,which is more than 11% higher than current mainstream video object segmentation methods based on deep learning.The average detection time for one frame image is 765 ms,which basically meets the needs of real-time detection.The developed rough-fine two-step TV region segmentation method can be integrated into the full-automatic leucorrhea sample analyzer.The experimental results of this research show that the developed autofocus method and formed elements segmentation and recognition method meet the requirements of assisting doctors in clinical examination in terms of detection accuracy and speed,and can be used in the full-automatic sample analyzer. |
PDF Full Text Request