Deep learning convolutional neural networks for morphological identification of field-caught Anopheles stephensi in Ethiopia, using smartphone images of specimens
4 affiliations
- 1Department of Biomedical Engineering, Center for Bioengineering Innovation and Design, Johns Hopkins University, Baltimore, MD USA
- 2Department of Biological Sciences, Eck Institute for Global Health, University of Notre Dame, Notre Dame, IN USA
- 3Department of Environmental Health Sciences and Technology, Jimma University, Jimma, Ethiopia
- 4Bloomberg School of Public Health, Johns Hopkins Malaria Research Institute, Johns Hopkins University, Baltimore, MD USA
Abstract
Background: Anopheles stephensi, an invasive malaria vector historically confined to South Asia and the Arabian Peninsula, has established reproducing populations across the Horn of Africa and continues to expand its range. Its adaptability to a range of environments, with its impact described most prominently in urban and peri-urban settings, and morphological similarity to native African anophelines, particularly An. gambiae s.l., makes reliable field identification difficult by field entomologists across much of Africa who have had limited prior exposure to the species. Existing molecular methods are impractical for routine field use, creating an urgent need for scalable, accessible identification tools to support surveillance programs responding to this emerging threat.
Methods: Three deep learning convolutional neural network (CNN) classifiers were developed and validated for detecting An. stephensi from photographic images of mosquito specimens collected from sentinel sites across Ethiopia and Uganda, verified for species identity by PCR or morphological identification. VectorCam, a standardized field imaging device, served as the primary imaging platform, with a supplementary handheld smartphone dataset collected using clip-on macro lenses (10X and 15X). An EfficientNet-B1 architecture was fine-tuned using transfer learning from a pretraining dataset of over 60,000 labeled photographs of field-collected mosquito specimens from Uganda. Three classifiers were developed: a VectorCam multiclass classifier distinguishing seven species classes including An. stephensi, a VectorCam binary classifier distinguishing An. stephensi from all other species, and a handheld smartphone binary classifier for photographic images taken using a handheld smartphone. All models were evaluated using 5-fold cross-validation and a majority voting ensemble of the trained classifiers.
Results: Against PCR and morphologically verified ground-truth identifications, the VectorCam multiclass classifier achieved an An. stephensi sensitivity of 98.74% and overall accuracy of 96.02%. The VectorCam binary classifier achieved an An. stephensi sensitivity of 97.48% and a specificity of 99.25%. The handheld binary classifier achieved an An. stephensi sensitivity of 91.09% and a specificity of 84.40%. Gradient-weighted Class Activation Mapping confirmed that classification decisions were driven by wing venation and leg morphology, consistent with established entomological identification criteria.
Conclusions: Deep learning image classification using VectorCam achieves high accuracy for field detection of An. stephensi, with the binary classifier demonstrating performance suitable for operational vector surveillance, providing a practical, scalable solution for decentralized An. stephensi surveillance across malaria-endemic regions of Africa, reducing reliance on molecular verification or specialist taxonomic training for field-level species identification.
Related
- Evaluating the intrinsic adulticidal properties of a few commercial wall paints on Anopheles stephensi mosquitoes
- Anopheles larval ecology and physicochemical characterization of larval habitats in Dire Dawa: an area colonized by Anopheles stephensi in Eastern Ethiopia.
- Passive Smartphone-Based Digital Phenotyping for the Identification of PTSD Symptoms: Initial Validation of a Digital Behavioral Marker
- Construction and validation of a unified 3D convolutional neural network framework for automated identification of cervical spinal cord injury based on CT imaging
- High-Throughput Morphological Profiling of Patient-Derived Human Liver Organoids for Hepatotoxicity Hazard Identification and Compound Prioritization
BibTeX
@article{mehta2026deep,
title = {{Deep learning convolutional neural networks for morphological identification of field-caught Anopheles stephensi in Ethiopia, using smartphone images of specimens}},
author = {Parthvi Mehta and Aravind Sunil Kumar and Atul Zacharias and Aryaman Shodhan and Deming Li and Sunny M. Patel and Marina Rincon Torroella and Joanne M. Cunningham and Diane Lovin and Biniam Lukas Chbute and Deneke Zewdu Gebremichael and Rama Chellappa and Delenasaw Yewhalaw and Jane M. Carlton and Neil F. Lobo and Soumyadipta Acharya},
journal = {Malaria Journal},
year = {2026},
month = sep,
doi = {10.1186/s12936-026-06122-5},
url = {https://doi.org/10.1186/s12936-026-06122-5},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Deep learning convolutional neural networks for morphological identification of field-caught Anopheles stephensi in Ethiopia, using smartphone images of specimens
AU - Parthvi Mehta
AU - Aravind Sunil Kumar
AU - Atul Zacharias
AU - Aryaman Shodhan
AU - Deming Li
AU - Sunny M. Patel
AU - Marina Rincon Torroella
AU - Joanne M. Cunningham
AU - Diane Lovin
AU - Biniam Lukas Chbute
AU - Deneke Zewdu Gebremichael
AU - Rama Chellappa
AU - Delenasaw Yewhalaw
AU - Jane M. Carlton
AU - Neil F. Lobo
AU - Soumyadipta Acharya
PY - 2026
DA - 2026/09/17
JO - Malaria Journal
DO - 10.1186/s12936-026-06122-5
UR - https://doi.org/10.1186/s12936-026-06122-5
AB - Background: Anopheles stephensi, an invasive malaria vector historically confined to South Asia and the Arabian Peninsula, has established reproducing populations across the Horn of Africa and continues to expand its range. Its adaptability to a range of environments, with its impact described most prominently in urban and peri-urban settings, and morphological similarity to native African anophelines, particularly An. gambiae s.l., makes reliable field identification difficult by field entomologists across much of Africa who have had limited prior exposure to the species. Existing molecular methods are impractical for routine field use, creating an urgent need for scalable, accessible identification tools to support surveillance programs responding to this emerging threat. Methods: Three deep learning convolutional neural network (CNN) classifiers were developed and validated for detecting An. stephensi from photographic images of mosquito specimens collected from sentinel sites across Ethiopia and Uganda, verified for species identity by PCR or morphological identification. VectorCam, a standardized field imaging device, served as the primary imaging platform, with a supplementary handheld smartphone dataset collected using clip-on macro lenses (10X and 15X). An EfficientNet-B1 architecture was fine-tuned using transfer learning from a pretraining dataset of over 60,000 labeled photographs of field-collected mosquito specimens from Uganda. Three classifiers were developed: a VectorCam multiclass classifier distinguishing seven species classes including An. stephensi, a VectorCam binary classifier distinguishing An. stephensi from all other species, and a handheld smartphone binary classifier for photographic images taken using a handheld smartphone. All models were evaluated using 5-fold cross-validation and a majority voting ensemble of the trained classifiers. Results: Against PCR and morphologically verified ground-truth identifications, the VectorCam multiclass classifier achieved an An. stephensi sensitivity of 98.74% and overall accuracy of 96.02%. The VectorCam binary classifier achieved an An. stephensi sensitivity of 97.48% and a specificity of 99.25%. The handheld binary classifier achieved an An. stephensi sensitivity of 91.09% and a specificity of 84.40%. Gradient-weighted Class Activation Mapping confirmed that classification decisions were driven by wing venation and leg morphology, consistent with established entomological identification criteria. Conclusions: Deep learning image classification using VectorCam achieves high accuracy for field detection of An. stephensi, with the binary classifier demonstrating performance suitable for operational vector surveillance, providing a practical, scalable solution for decentralized An. stephensi surveillance across malaria-endemic regions of Africa, reducing reliance on molecular verification or specialist taxonomic training for field-level species identification.
SN - 1475-2875
ER -