Abstract
Automated detection of malaria trophozoites in Giemsa-stained blood smears from Africa is primarily constrained by false-positive (FP) detections rather than inadequate sensitivity. In resource-limited settings, each FP increases the burden of confirmatory microscopy, placing additional strain on already scarce laboratory resources. Using the Lacuna Malaria dataset, comprising samples from Uganda and Ghana with 15{,}838 parasites and 7{,}004 white blood cell (WBC) annotations, this study investigated the factors contributing to FPs under a standardized evaluation protocol (IoU = 0.5, confidence = 0.20). A spatial null model demonstrated that FPs were significantly less likely to occur near WBCs than expected by chance ($z = -2.55$), disproving the hypothesis that WBCs are the primary source of erroneous parasite detections. Instead, FPs were associated with image blur and a slight bias toward larger object sizes. Several strategies for reducing FPs were evaluated against a two-class baseline of 892 false detections. A saturated log-area loss reduced FPs by 7\%, hard-negative mining achieved a 27\% reduction, and test-time augmentation combined with weighted box fusion produced the largest improvement, lowering FPs from 892 to 542 while increasing the F1 score from 0.624 to 0.719. These improvements were consistently observed across YOLO26x, RT-DETR-X, and YOLOv5m6 models, all converging on an F1 score of approximately 0.72. Analysis of the remaining FPs revealed that most occurred in regions consistently identified as parasites by all six detectors but lacking ground-truth annotations, suggesting the presence of unlabelled parasites rather than true detection errors. Overall, the findings indicate that improving annotation quality and employing appropriately calibrated, efficiently sized models yields greater performance gains than simply increasing model complexity, particularly in resource-constrained clinical settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_042.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_042_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=zLSeR9JQVP
BibTex
@InProceedings{OghSup_FalsePositive_MICCAISAT2026,
author = { Oghenewoakpo, Supreme Onowoakpo AND Supreme, Mercy Ajoke AND Kukudabi, Seth Kwabena Kyei AND Yakubu, Kanyiri Ahmed AND Yinsuu, Gideon AND Buernorkie Agbugblah, Doreen AND Beckley, Oserebameh Augustine AND Agwu, Mary AND Zhang, Dong AND Raymond, Confidence AND Iorumbur, Aondona Moses AND Mumuni, Abdul Nashirudeen AND Opara, Chidera AND Adewole, Maruf AND Tigbee, Isaac},
title = { { False-Positive Reduction in Automated Malaria Parasite Detection from African Blood Smears: A Falsification-to-Mitigation Study } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}