Research On Malaria Detection And Classification Based On Deep Learning

Malaria is a serious tropical infectious disease that still affects hundreds of millions of people.In 2021,due to the global trend of COVID-19 pandemic,public health services in many malaria-endemic areas were forced to stop,which further intensified the global malaria epidemic in 2021,with a total of about 247 million malaria cases and619,000 deaths.Microscopy is an authoritative malaria detection method,but in resource-poor areas,there is a shortage of detection personnel,unable to obtain accurate detection results,which seriously affects the follow-up intervention and treatment of malaria.Aiming at this problem,this paper applies computer-aided detection and combines deep learning with malaria detection tasks to achieve accurate classification of malaria and accurate detection of Plasmodium.The main research results of this paper are as follows:(1)In the study on the classification of infected and uninfected malaria thin blood smear images,it is found that the ResNet-50 model has the problems of long training time and low training accuracy.To solve this problem,the ResNet-40+CBAM network is proposed in this paper.CBAM was introduced into ResNet-40 network model,and CBAM modules were embedded in different positions of ResNet-40 network structure to verify the feature enhancement of attention mechanism.The proposed model was trained on the Malaria data set of thin blood smear images for malaria and compared with other models.Experimental results show that the performance of ResNet-C-40 model is the best,the accuracy of 98.42%,compared with VGG-16,ResNet-50,ResNet-40,ResNet-C-Res-40,the accuracy of 1.5%～4.0% higher.The training time of the model is 0.9h,which is much smaller than that of other models.The size of the model is only 18.3MB,making it easy to deploy on mobile devices.Based on the above experimental results,the effectiveness of the proposed method is also proved.(2)In the detection of malaria parasites in thick blood smear images,it is found that the accuracy of detection is low in the Faster-RCNN network model.To solve this problem,the Mask-RCNN-D algorithm is proposed in this paper.According to the characteristics of the model,the original data set was processed before the training.Firstly,the malaria parasite was cut out according to the XML annotation file,and the picture size was defined as 64*64,that is,the minimum trainable size of Mask-RCNN network.After the cutting,more than 40,000 small Plasmodium images were re-labeled to create a new Plasmodium data set.Mask-RCNN-D algorithm introduces Dilated Convolution to better preserve deep and shallow image features,so as to improve detection accuracy.When the model is trained on the Plasmodium data set,Mask-RCNN-D(dr=2)has the highest detection accuracy,mAP is 93.75%.The training results on Mask-RCNN-D(dr=2)and Mask-RCNN are both higher than those of the Faster-RCNN model.Increased by 2.33% and 3.77%.The above experimental results prove the effectiveness of the proposed algorithm.