Malaria Cell Detection Based On Improved YOLO V3

Malaria is a mosquito-borne blood disease caused by the transmission of Plasmodium by the bite of female Anopheles mosquitoes.As one of the parasitic diseases that seriously endanger human health,Malaria is widely distributed in more than 90 countries and regions in tropical,subtropical and temperate zones.Accurate detection of malaria-infected cells is a prerequisite for subsequent treatment.Although microscopic examination is the "gold standard" for malaria detection,the test results are accurate but time-consuming and the reliability of the test results depends on the professional level of the test personnel.The introduction of deep learning to assist or even replace manual microscopy can not only ensure the accuracy of rapid diagnosis,but also save a lot of funds for diagnosis and control of malaria.The YOLO V3 target detection algorithm has a good effect on the detection of dense small targets in the picture,but there is still room for improvement in the malaria detection task.In this thesis,YOLO V3 has been improved in many aspects to realize the integration of high-precision malaria cell localization and classification to assist the clinical diagnosis of malaria.Improvements can be listed as follows:1)Optimization of cell localization.The IOU positioning loss function originally used in YOLO V3 has the problems of inaccurate positioning effect and slow convergence speed.This thesis introduces the GIOU and DIOU localization loss functions.The results of the comparative experiments show that the localization effects of the GIOU loss function and the DIOU loss function are better than the IOU loss function,and the DIOU loss function performs better.2)Optimization of cell classification.The optimization of cell classification includes the optimization of class classification and confidence classification:for the problem that the number of normal cells in the malaria thin blood smear is large and the number of infected cells is small（that is,the classification is severely imbalanced）,label regularization is introduced in the classification process.The cell category is smoothed,and weights are set for infected cells,so that the model pays more attention to the correct classification of infected cells and improves the problem of category imbalance;then the QFL loss function is introduced to reduce the possibility of category classification and confidence classification errors at the same time sex.3)Optimization of output link.When the NMS non-maximum suppression method selects candidate boxes,it will lose part of the information output by the model,which is ineffective for the selection of prediction boxes.In this thesis,the test-time augmentation（TTA）method and the non-maximum suppression method of weighted boxes fusion（WBF）are used to adjust the predicted frame output by YOLO V3 to obtain better output results.Through the three aspects of improvement mentioned above,the detection accuracy of YOLO V3 has been effectively improved.The detection accuracy of the optimal single model reaches 82.82%,which is 6.11%higher than before.Finally,by using the idea of ensemble learning,multiple training models were fused by WBF method,which effectively reduced the phenomenon of missing detection,and the detection accuracy of the optimal combination model reached 86.05%.The results show that the improved YOLO V3 proposed in this thesis is feasible and effective,and can better realize the detection of malaria thin blood smear cells.