Study On Deep Learning Based Automated Cell Segmentation And Classification Of Malaria Microscope Image

Malaria is one of the major health threats in the world.According to the latest report of the World Health Organization,400000 people die from malaria every year.Timely and accurate malaria diagnosis is the key of malaria follow-up treatment and prevention and control.However,at present,malaria diagnosis mainly depends on experienced professional technicians,and each patient diagnosis takes about 30 minutes,so using artificial intelligence technology to automatically segmentation and classification all kinds of cells on the malaria microscope image quickly and accurately has important clinical significance and application value.This paper proposes a cell segmentation and classification algorithm for malaria microscope image based on deep learning,which can achieve high-precision cell automatic segmentation and classification,and achieve the integration of segmentation and classification,so as to assist the clinical diagnosis of malaria.The main contributions of this paper are as follows.(1)A multi-class cell segmentation algorithm based on RU-Net and a cell classification algorithm based on S-Faster R-CNN are proposed.RU-Net algorithm realizes multi-class segmentation,and combines residual network with U-Net network,deepens the network depth of U-Net and improves the segmentation accuracy of cell.RU-Net algorithm obtains 87.90% Dice coefficient on MIT malaria dataset;S-Faster R-CNN algorithm realizes multi-label classification on malaria microscope image.Because there are many overlapped cells in the malaria microscope image,the algorithm introduces Soft Non-maximum Suppression to deal with the overlapped detection boxes and reduce the missed detection rate of overlapped cells.S-Faster R-CNN is also tested on the MIT-Malaria dataset,and the mean Average Precision is used as the evaluation index,S-Faster R-CNN achieved 77.58% mean Average Precision.(2)Based on FD-Mask R-CNN,an algorithm of cell segmentation and classification is proposed to realize the integration of cells segmentation and classification.On the basis of Mask R-CNN,two improvements have been made: first,in order to improve the segmentation precision,the algorithm introduced the dilated convolution to increase the convolution receptive field and retain more spatial structure of image.Second,in the malaria microscope image,due to the large number of normal red blood cells and the small number of plasmodium cells,there is a serious class imbalance.In order to solve this problem,we add modulation factor in classification loss function to strengthen the attention of the difficult classification samples,improve the class imbalance problem,improve the classification average precision.The experimental results based on the MIT-Malaria dataset showed that FD-Mask R-CNN obtained 95.74% Dice coefficient in segmentation,and 88.18% mean Average Precision in classification.Compared with RU-Net and S-Faster R-CNN,it has a certain improvement in segmentation and classification precision.The experimental results above show that the proposed algorithm in this paper is feasible and effective,and can be extended to the segmentation and classification of other cell images,with a certain generalization and potential clinical application value.