关键词: 疟原虫检测, 目标检测模型, YOLOv8模型, 深度学习, 血涂片

Abstract:

Objective To develop and establish an artificial intelligence (AI) platform assisting malaria parasites detection, and the applied effect was evaluated. Methods Blood smears were collected from current cases of overseas imported malaria in Wuhan City from 2008 to 2022. Microscopic images were captured with an automatic imaging collection system, and annotated by professional microscopists to create a dataset of thin blood smear images. A detection model was employed to detect Plasmodium, white blood cells (WBCs) and red blood cells (RBCs) to deduce negativity, positivity, infection rate and infected parasite species, and the results were displayed in the application interface. Based on numbers of true positive, true negative, false positive and false negative samples, the accuracy, precision, recall, and F1 score were estimated to evaluate the performance of the model. Microscopic examination procedures were mimicked according to the technical assessment method of malaria microscopists, and 10 blood smears were randomly selected to score the model comprehensive results. Results The thin blood smears image datasets included 3 704 images that were annotated by microscopist, including 7 835 Plasmodium annotations and 786 WBC annotations. Results from the YOLOv8 confusion matrix showed an overall accuracy of 85.5%, precision of 98.5%, recall of 93.3%, F1 score of 95.8%, and an mAP@0.50 score of 96.2%. The model had a high accuracy for detection of P. vivax (96.9%) and P. malariae (93.8%), and a low accuracy for detection of P. falciparum (87.2%) and WBCs (87.8%). Except an 89.7% F1 score for detection of P. falciparum, the model had F1 scores of over 90.0% for detection of other Plasmodium species and WBCs. The recall of the model for detection of 4 Plasmodium species and WBCs, was the highest recall for detection of P. malariae (94.9%), followed by WBC (94.2%), P. vivax (94.1%), P. ovale (93.6%) and P. falciparum (92.4%). A total of 64 targets were falsely identified, with a false detection rate of 6.7%. The correction rates of the platform and professional microscopist were both 100%, and the clinical microscopy score was 90 points, with no false positives or negatives identified; however, there were false detections in identification of malaria parasite species. Conclusion The platform is effective to identify and locate malaria parasites, WBCs and RBCs, and its scoring results are approaching to the microscopist. Due to a high detection capability, the model is feasible for auxiliary diagnosis of clinical microscopy of malaria parasites.

Key words: Plasmodium detection, Object detection model, YOLOv8 model, Deep learning, Blood smear