A feature pyramid classifier for Trypanosoma cruzi in smartphone blood smear microscopy

BACKGROUND: Chagas disease, caused by Trypanosoma cruzi (T. cruzi), remains a public health challenge in endemic regions, with an estimated 70 million people at risk. "Gold standard" diagnostics rely primarily on serologic or molecular techniques that require specialized equipment often inaccessible in low-resource settings. Light microscopy is a low-cost diagnostic method common across such regions when specialized parasitology expertise is available. Computer vision screening tools that run on smartphone-acquired microscopy may improve accessibility to high-quality accurate diagnoses. METHODOLOGY/PRINCIPAL FINDINGS: We developed a custom feature pyramid network (FPN) for image level classification of T. cruzi in thin blood smear quadrants from the publicly available dataset of Morais et al. (33 slides, 704 smartphone acquired fields). We partitioned the data at the level of microscope fields, assigning fields 0600-0704 to the test set in accordance with the held out block originally defined by Morais et al. Training used AdamW with binary cross-entropy loss. Across five independent random seeds, test set AUC was tightly concentrated at 0.934 ± 0.019 (range 0.916-0.961). The best of the five runs achieved 89.5% accuracy, 95.7% precision, 88.2% sensitivity, 92.2% specificity, F1 91.8% and AUC 0.961 at the default 0.5 decision threshold. A validation selected threshold of 0.355 lifts that run's test recall to 92.8% and F1 to 93.8%. To verify the model attends to parasite morphology rather than to spurious image features, we applied gradient-weighted class activation maps (Grad-CAM), an explainable AI technique that localizes where the model looks. Quantified against Morais's parasite centroid annotations, the peak attention fell within 150 pixels of an annotated parasite on 88.9% of true positives. A blinded re-audit of the model's test set disagreements by three independent epidemiologists further identified 6 of 11 (54.5%) apparent false positives and 3 of 33 (9.1%) apparent false negatives as annotation errors rather than model mistakes. CONCLUSIONS/SIGNIFICANCE: A compact CNN/FPN architecture is effective for image level T. cruzi classification on smartphone microscopy and additionally surfaces candidate annotation errors for targeted expert verification. These results motivate prospective validation on multi-laboratory, multi-geography datasets and extension to object level detection and counting.

Authors

Institutions

Publication Details

Journal
PLoS neglected tropical diseases
Published
2026-11-17
DOI
https://doi.org/10.1371/journal.pntd.0014882
Primary Topic
Digital Imaging for Blood Diseases
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A feature pyramid classifier for Trypanosoma cruzi in smartphone blood smear microscopy

Melissa S. Nolan, Mary K. Lynn, Homayoun Valafar, Kyndall C. Dye‐Braumuller et al.
1 links
PLoS neglected tropical diseases
Digital Imaging for Blood Diseases
article

A feature pyramid classifier for Trypanosoma cruzi in smartphone blood smear microscopy

Melissa S. Nolan, Mary K. Lynn, Homayoun Valafar, Kyndall C. Dye‐Braumuller, Lídia Gual-González, Christopher Lee
article en

Abstract

BACKGROUND: Chagas disease, caused by Trypanosoma cruzi (T. cruzi), remains a public health challenge in endemic regions, with an estimated 70 million people at risk. "Gold standard" diagnostics rely primarily on serologic or molecular techniques that require specialized equipment often inaccessible in low-resource settings. Light microscopy is a low-cost diagnostic method common across such regions when specialized parasitology expertise is available. Computer vision screening tools that run on smartphone-acquired microscopy may improve accessibility to high-quality accurate diagnoses. METHODOLOGY/PRINCIPAL FINDINGS: We developed a custom feature pyramid network (FPN) for image level classification of T. cruzi in thin blood smear quadrants from the publicly available dataset of Morais et al. (33 slides, 704 smartphone acquired fields). We partitioned the data at the level of microscope fields, assigning fields 0600-0704 to the test set in accordance with the held out block originally defined by Morais et al. Training used AdamW with binary cross-entropy loss. Across five independent random seeds, test set AUC was tightly concentrated at 0.934 ± 0.019 (range 0.916-0.961). The best of the five runs achieved 89.5% accuracy, 95.7% precision, 88.2% sensitivity, 92.2% specificity, F1 91.8% and AUC 0.961 at the default 0.5 decision threshold. A validation selected threshold of 0.355 lifts that run's test recall to 92.8% and F1 to 93.8%. To verify the model attends to parasite morphology rather than to spurious image features, we applied gradient-weighted class activation maps (Grad-CAM), an explainable AI technique that localizes where the model looks. Quantified against Morais's parasite centroid annotations, the peak attention fell within 150 pixels of an annotated parasite on 88.9% of true positives. A blinded re-audit of the model's test set disagreements by three independent epidemiologists further identified 6 of 11 (54.5%) apparent false positives and 3 of 33 (9.1%) apparent false negatives as annotation errors rather than model mistakes. CONCLUSIONS/SIGNIFICANCE: A compact CNN/FPN architecture is effective for image level T. cruzi classification on smartphone microscopy and additionally surfaces candidate annotation errors for targeted expert verification. These results motivate prospective validation on multi-laboratory, multi-geography datasets and extension to object level detection and counting.

PLoS neglected tropical diseasesVol. 20(11)
University of South Carolina (US), Clemson University (US)
National Science Foundation, National Institute of General Medical Sciences
Good health and well-being
Digital Imaging for Blood Diseases
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon