Research And Application On Intelligent Classification About Medical Image

Recent years have seen the booming research in computer techniques towards a fully developed system.This has contributed significantly to the clinical diagnosis and assisted detection within medical profession.Cervical cancer is the only kind of malignant neoplasms that is with clear cause,preventability,and right available treatment during early and middle stage.In time discovery of the cervical cancer will indeed help increase the cure rate and decrease its morbidity.Currently the development of the computer-aided cervical cancer detection system is still at its early stage.This thesis focuses on the classification recognition research of cervical cancer cell images.The task is achieved with the help of pioneering works in this area,technologies from relevant disciplines such as mage segmentation,machine learning and cytopathology knowledge.Under the guidance of pathologists,the cell images are pre-processed,segmented,feature-extracted,which eventually lead to the classification recognition of cervical cancer cell.Major contributions are listed below:1.An improved Otsu double threshold method is used for the rough division of cervical cancer cell images.The areas of interest in the images are extracted and adhesion discrimination were performed.This is followed by precise division of the images within areas of interest based on Chan-Vase model,where the images of cell body,cytoplasm,nucleus are obtained.Single cells and adherent cells are marked respectively.Instead of forced segmenting using morphological methods,adherent cells are segmented as a whole,which retains the morphological integrity of cells to a maximum.2.Differential extraction for the characteristic parameters of the cells is performed based on the precise division of the cells.This work focuses on the images of cells that are among single,with cytoplasmic adhesion found in single nucleus and with cytoplasmic adhesion found in both nucleus and nuclei.This work also features pioneering researches and Knowledge of cervical cytopathology,and is under the guidance of pathologists.The parameters obtained are stored as data in SQL Server database.3.The cervical cell classifier is designed based on SVM classifier principle owing to its good performance.The programme coding is based on Visio Studio software platform and LIBSVM open source software package,where C++ is selected as the programming language.Separate classifiers are designed dedicatedly for single cell,cell with cytoplasmic adhesion found in single nucleus and cell with cytoplasmic adhesion found in both nucleus and nuclei.After comparison,the classifier design utilises the decision tree method with relatively less classifiers constructed,while RBF Radial Basis Function is selected as the kernel function of the classifier model.The data stored in characteristic parameters table from the SQL Server is used as sample sets,the grid search algorithm is used for determining the parameters of SVM classifier,and the K-fold cross validation is used for training the classifier model,which eventually leads to the optimised classifier.The results show that the cervical cancer cell images classifier designed following the above mentioned steps behaves of expectation for not only single cervical cell,but also cervical cells stuck together.This thesis may further contribute as a reference for future development of the computer-aided cervical cancer detection system.