Source-linked AI summary
Investigating White Blood Cells as a Source of False-Positive Malaria Parasite Detection in African Blood-Smear Images
Samuel A. Adeniji, Goodness C. Obasi, Chris-Victor Ntwali, Aondana M. Iorumbur, Confidence Raymond, Lowami Uwimana, Ahmed Tahiru Issah
TL;DR
False-positive malaria detections can prompt unnecessary treatment where confirmatory testing is unavailable, and WBCs resemble early-stage parasites. This study compares parasite-only and parasite-plus-WBC detectors and tests whether false positives cluster near WBCs; the analyses instead identify predominantly background detections and spatial repulsion.
Problem
Whether visually similar WBCs account for parasite-detector false positives matters because false alarms can trigger unnecessary treatment where confirmatory testing is unavailable.
Source: §1 · p. 2Method
The study compares controlled YOLOv12s models trained with parasite-only versus parasite-and-WBC labels and applies seven spatial and statistical analyses to false-positive locations.
Source: §1 · p. 3, §2.1 · p. 3, §2.4 · p. 4Results
95% of false positives are background detections, none are WBC class confusions, and spatial tests show repulsion between false positives and WBCs; Model B’s advantage is attributed to multi-task representations.
Source: §4 · pp. 7–8Takeaways & Limitations
The findings point away from WBC confusion and toward background noise, Giemsa debris, slide-preparation variability, and possibly unlabeled early-stage parasites as false-positive sources.
Source: §4 · pp. 7–8
Abstract
from arXiv · showhide
White blood cells (WBCs) present on every Giemsa-stained thick blood smear share visual properties with early-stage Plasmodium falciparum ring-form trophozoites: small size, round morphology, and intense purple staining. They are a plausible but untested source of false positives in parasite-only detectors. We trained two YOLOv12s models on the Lacuna Malaria Detection dataset (8,000 images from Uganda and Ghana): Model A with parasite labels only, and Model B with both parasite and WBC labels. Seven independent spatial and statistical analyses tested whether false positive (FP) predictions cluster near WBC locations. All seven refute the hypothesis. In both models, 95% of FPs are pure background detections (IoU below 0.10 against any ground-truth box); zero are WBC class confusions. Ripley's Cross-K analysis shows spatial repulsion between FP centroids and WBC positions at every radius tested. Model B outperforms Model A overall (mAP50 0.859 vs. 0.755), and the advantage is uniform across all WBC-proximity bands, pointing to multi-task representation learning rather than WBC suppression as the cause. False positives arise from Giemsa stain debris and preparation artifacts. Effective mitigation requires staining artifact augmentation and annotation of unannotated early-stage ring forms rather than WBC labeling alone.
1 Introduction
False positives in malaria smear detection can lead to unnecessary treatment, while WBCs resemble early-stage parasites and have been proposed as a source of errors. This study tests that hypothesis using controlled parasite-only and parasite-plus-WBC detectors and spatial analyses.
1 Introduction
- Automated detection is motivated by microscopy’s staffing, observer-variability, and infrastructure constraints, while aggregate mAP can obscure false-positive burden.Prior work found that sensitivity gains can accompany higher false-positive rates.Source: §1 · p. 2
- False positives can prompt unnecessary antimalarial treatment, especially where confirmatory testing is unavailable.Thick-smear preparation also introduces debris, stain precipitate, and overlapping material that can trigger spurious detections.Source: §1 · p. 2
- No prior study had directly tested whether parasite-detector false positives are spatially linked to WBC locations.Although WBCs share small size, round morphology, and intense purple staining with early-stage ring forms, their role as a false-positive source remained untested.Source: §1 · pp. 2–3
- The study compares otherwise controlled YOLOv12s models trained with parasite-only versus parasite-and-WBC labels on 8,000 Uganda and Ghana images.Precision is the primary metric, and seven spatial and statistical analyses test whether false positives cluster near WBCs.Source: §1 · p. 3
2 Methodology
The methodology compares two YOLOv12s detectors on annotated thick-smear images, differing in whether WBCs are labeled, and evaluates detection errors and their spatial relation to WBCs.
2.1 Dataset and Experimental Design
- The Lacuna Malaria Detection Challenge dataset provides digitized thick Giemsa-stained blood-smear images at 2,048 × 2,048 pixels; its held-out test set contains 412 images and 2,374 parasite ground-truth instances.Model B adds WBC as a second class, with 1,195 WBC test instances.Source: §2.1 · p. 3
2.2 YOLOv12s Architecture and Training
- Both detectors use YOLOv12s with identical training hyperparameters, so the annotation strategy is the sole intended difference.Model B’s WBC centroid predictions also supply locations for the spatial false-positive analysis.Source: §2.2 · p. 3
2.3 False Positive Extraction and TIDE Analysis
- Predictions were extracted at confidence ≥0.10 and greedily matched in descending confidence to ground truth at IoU ≥0.50; TIDE partitions false positives and false negatives into six mutually exclusive error types.Source: §2.3 · pp. 3–4
2.4 Spatial Analysis and Statistical Tests
- Seven analyses assess WBC proximity through centroid distances, proximity bands, spatial statistics, density surfaces, overdispersion, and clustering.Paired bootstrap, Wilcoxon, and BH-corrected stratified McNemar tests compare models; temperature scaling and multivariate isotonic regression assess calibration.Source: §2.4 · p. 4
3 Results
Model B leads Model A on aggregate detection performance, but the analyses find that false positives are predominantly background detections rather than WBC confusions. Spatial tests reject WBC proximity as their source, while model comparisons and calibration characterize the broader performance differences.
3.1 Baseline Detection Performance
- mAP50 0.859 vs. 0.755: Model B leads Model A overall, though parasite mAP50 is 0.748 vs. 0.755.The WBC class has mAP50 = 0.970 and recall = 0.965; the aggregate lead therefore coexists with a small parasite-class trade-off.Source: §3.1 · pp. 4–5, Table 1 · p. 5
3.2 False Positive Characterization
- 95.1% of Model A and 95.3% of Model B false positives are background detections, with zero classification errors in either model.No Model B parasite prediction overlaps a WBC ground-truth box at IoU ≥0.50.Source: §3.2 · p. 5, Table 2 · p. 5
3.3 Spatial Analysis
- Over 79% of false positives lie more than 300 px from WBCs, and Ripley’s Cross-K shows repulsion at every tested radius from 10 to 300 px.In Model B, the false-positive rate rises from 17.6% at 0–30 px to 27.8% beyond 300 px.Source: §3.3 · p. 6, Table 3 · p. 6
- False-positive counts are highly overdispersed by image, with negative-binomial models preferred over Poisson by delta-AIC ≈800 for both detectors.The analysis links this image dependence to slide-preparation variability rather than WBC density.Source: §3.3 · p. 6, Table 4 · p. 6
3.4 Statistical Model Comparison
- Model B’s mAP50 advantage is +0.0234 (95% BCa CI: [+0.006, +0.042]), and it is significantly better in all four powered WBC-proximity bands.Model B correctly detects 136 annotations missed by Model A, versus 38 in the reverse direction; the distal >300 px band also favors Model B (p = 0.003).Source: §3.4 · p. 7, Table 5 · p. 7
3.5 Confidence Calibration
- Multivariate isotonic regression reduces Model B parasite-class ECE from 0.123 to 0.023, an 81% reduction.Box height (+0.291) and width (−0.285) are the dominant calibration features.Source: §3.5 · p. 7
4 Discussion
Across seven analyses, false positives are not caused by confusion with WBCs. Model B’s advantage is attributed to multi-task representation learning, not WBC-error suppression.
4 Discussion
- 95% of false positives are background detections, with zero WBC class-confusion errors; FP locations are spatially repelled from WBCs.False positives instead trace to background noise, Giemsa stain debris, lysed-cell fragments, and possibly unlabeled early-stage parasites.Source: §4 · p. 7
- False-positive patterns track slide-preparation quality rather than WBC density.The reported overdispersion analysis supports this distinction.Source: §4 · pp. 7–8
- Model B’s advantage over Model A reflects better learned representations through multi-task training, not suppression of WBC-related errors.Source: §4 · p. 8
5 Impact on Low-Resource Settings
False positives remain a serious clinical concern where confirmatory testing is unavailable. Preparation variability calls for broader training coverage, while WBC labels offer a practical upgrade with no added inference cost.
5 Impact on Low-Resource Settings
- At a realistic operating threshold, Model A’s precision is 51%, so roughly half of its flagged parasites are wrong.The passage identifies this as a serious problem where confirmatory testing is unavailable.Source: §5 · p. 8
- Because false positives stem from preparation variability rather than WBC proximity, the proposed fix is broader staining-condition coverage, not WBC labels alone.Source: §5 · p. 8
- Adding WBC labels improves precision from 51% to 57% and calibration, with no added compute or inference cost.Microscopists already count WBCs as part of standard practice, making this a practical upgrade path for resource-constrained clinics.Source: §5 · p. 8
6 Conclusion
This study tests whether WBCs cause false positives in parasite detection using a controlled comparison of two YOLOv12s models and seven analyses. The analyses reject WBC confusion as the explanation, while the model with parasite and WBC labels achieves higher mAP50.
6 Conclusion
- 95% of false positives are background detections unrelated to any cell, with none classified as WBC confusions and FP locations repelled from WBCs.Model B’s higher mAP50 (0.859 vs. 0.755) is attributed to richer shared feature learning from multi-task training rather than WBC suppression.Source: §6 · p. 8
- The controlled comparison uses two YOLOv12s detectors trained on 8,000 Lacuna images: parasite-only Model A and parasite-plus-WBC Model B.Source: §6 · p. 8
- The study addresses a suspected but previously untested source of false positives: visually similar WBCs in Giemsa-stained smears.False positives carry direct clinical cost where confirmatory testing is unavailable.Source: §6 · p. 8
- The proposed practical response shifts from WBC annotation alone toward staining-artifact augmentation, broader preparation-condition coverage, and post-hoc confidence calibration.The passage states these add no extra inference cost.Source: §6 · p. 8
- Future work will test the framework on other suspected confounders and whether overdispersion generalizes to smear images from additional countries.Named confounders include unannotated early-stage parasites and platelet clumps.Source: §6 · p. 8