Research On Automatic Identification Of Fungi Of Formed Element In Leucorrhea

Leucorrhea microscopic examination is a routine item in obstetrics and gynecology examination.The indicator of fungi is an important criterion for the examination.If fungi in the leucorrhea is excessive,the woman is most likely to suffer from colpitis mycotica.However,at present,many small and medium-sized hospitals still rely on artificial microscopy for the detection of fungi in leucorrhea.Because there is no quantitative standard,doctors observe the fungi in the leucorrhea according to the size and shape,and fungi are identified by the years of experience of the doctors.Then the diagnosis is qualitatively judged based on the approximate number of observed fungi.Therefore,the result of artificial microscopy is easily influenced by external conditions,resulting in inaccurate diagnosis of the disease.In this dissertation,the microscope is used to obtain the leucorrhea microscopic images.Convolutional Neural Network,Histogram of Gradient algorithm,Principal Component Analysis algorithm and Support Vector Machine have been used to detect and identify the fungi in the leucorrhea microscopic images.The main works of this dissertation are as follows:Firstly,after obtaining clear leucorrhea microscopic images,these images need to be pre-processed,including five steps: gray scale process,morphological operation,image segmentation,screening and making a training dataset.The image segmentation contains the OTSU algorithm and screening use the connected component labeling algorithm.After the pre-process step is completed,a training images dataset can be obtained.Secondly,a convolutional neural network should be trained using the training dataset and this is the pre-training step.Then a complete convolutional neural network is obtained.For every image,the feature of gradients are extracted from each feature maps by using Histogram of Gradient algorithm and all the features of gradients can be integrated into a feature vector.The dimension of the feature vector should be reduced by using the Principal Component Analysis algorithm.At last,these reduced feature vectors are put into the Support Vector Machine for training and a model for identifying fungi can be obtained.Thirdly,500 leucorrhea microscopic images which contain fungi have been identified using the recognition model.These images contain 4792 fungi in total by artificial statistics.Using this algorithm,4533 fungi have been accurately identified,359 fungi have been missed,and 206 fungi have been wrongly detected.The recognition rate is 94.6%,the missed rate is 5.4% and the false rate is 4.3%.Finally,in this dissertation,an algorithm without Convolutional Neural Network is used to compare with this proposed algorithm.Compared with the proposed algorithm,this algorithm will miss a lot of information,because this algorithm detect fungi only based on the features in the original image,but the proposed algorithm extracts the features of all feature maps at the same time.Using this algorithm and the same 500 images,4792 fungi in total,only 4059 fungi have been accurately identified,but 733 fungi have been missed and 575 fungi have been wrongly detected.The recognition rate is 84.7%,the missed rate is 15.3% and the false rate is 12.0%.The proposed algorithm in this dissertation has greatly improved both in the missed detection rate and the false detection rate.In this dissertation,the automatic identification of fungi algorithm is proposed and applied to the leucorrhea microscopic examination,which has not been reported in China.By this algorithm,the recognition rate is 94.6% and the results show that the proposed algorithm in this dissertation not only meets the hospital's requirement for the detection of fungi in leucorrhea microscopic examination,but also has better results comparing to other methods in automatic identification of fungi.