Assessment of Expert-Level Automated Detection of Plasmodium falciparum in Digitized Thin Blood Smear Images.

This diagnostic study assesses an expert-level detection algorithm for Plasmodium falciparum, a bacteria that causes malaria, using a publicly available benchmark image data set. Question Can deep learning be used to develop an automated malaria detection algorithm? Findings In this diagnostic study that used a 1-stage deep learning framework and benchmark data sets, the malaria detection algorithm achieved expert-level performance in detecting Plasmodium falciparum in thin blood smear images. The comparable performance between the algorithm and human experts was confirmed by a clinical validation study at the cell level and the image level. Meaning The findings suggest that a clinically validated expert-level malaria detection algorithm could be used to accelerate the development of clinically applicable automated malaria diagnostics. Importance Decades of effort have been devoted to establishing an automated microscopic diagnosis of malaria, but there are challenges in achieving expert-level performance in real-world clinical settings because publicly available annotated data for benchmark and validation are required. Objective To assess an expert-level malaria detection algorithm using a publicly available benchmark image data set. Design, Setting, and Participants In this diagnostic study, clinically validated malaria image data sets, the Taiwan Images for Malaria Eradication (TIME), were created by digitizing thin blood smears acquired from patients with malaria selected from the biobank of the Taiwan Centers for Disease Control from January 1, 2003, to December 31, 2018. These smear images were annotated by 4 clinical laboratory scientists who worked in medical centers in Taiwan and trained for malaria microscopic diagnosis at the national reference laboratory of the Taiwan Centers for Disease Control. With TIME, a convolutional neural network-based object detection algorithm was developed for identification of malaria-infected red blood cells. A diagnostic challenge using another independent data set within TIME was performed to compare the algorithm performance against that of human experts as clinical validation. Main Outcomes and Measures Performance on detecting Plasmodium falciparum-infected blood cells was measured by average precision, and performance on detecting P falciparum infection at the image level was measured using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Results The TIME data sets contained 8145 images of 36 blood smears from patients with suspected malaria (30 P falciparum-positive and 6 P falciparum-negative smears) that had reliable annotations. For clinical validation, the average precision was 0.885 for detecting P falciparum-infected blood cells and 0.838 for ring form. For detecting P falciparum infection on blood smear images, the algorithm had expert-level performance (sensitivity, 0.995; specificity, 0.900; AUC, 0.997 [95% CI, 0.993-0.999]), especially in detecting ring form (sensitivity, 0.968; specificity, 0.960; AUC, 0.995 [95% CI, 0.990-0.998]) compared with experienced microscopists (mean sensitivity, 0.995 [95% CI, 0.993-0.998]; mean specificity, 0.955 [95% CI, 0.885-1.000]). Conclusions and Relevance The findings suggest that a clinically validated expert-level malaria detection algorithm can be developed by using reliable data sets.