Dados do Trabalho

Título

Automatic identification of Trypanosoma cruzi vectors using deep learning techniques

Introdução

Trypanosoma cruzi is a flagellate parasite and the etiologic agent of Chagas disease. Its vectors are hematophagous insects of the family Reduviidae, subfamily Triatominae, in which 159 species are recognized. Many of these species are confused with other insects, especially non-haematophagous true bugs (Heteroptera). This similarity emphasizes the need for the involvement of experts adept at identifying Triatomines through image analysis. Therefore, an application of deep learning techniques for automatic classification of triatomines could be used as a triage system in the identification of species sent by the population via mobile phone, providing a fast and efficient preliminary identification, reducing the time needed by the experts.

Objetivo (s)

Development of a system for the automatic identification of Chagas disease vectors, using deep learning techniques to review photos submitted by the community, and integration of this system into the Fiocruz Chagas Disease Portal.

Material e Métodos

The images were collected from the Fiocruz website, the GeoVin Project, and the (National and International Reference Laboratory in Taxonomy of Triatomines) LNIRTT image bank, and divided into 2 classes: 330 images of triatomines and 375 images of other insects. A deep learning algorithm based on the ResNet18 architecture of CNNs was used to classify triatomines. After training, the model was validated on images from the Fiocruz portal using the SoftMax activation function to provide classification percentages. The routine was implemented in Python on Google Colaboratory.

Resultados e Conclusão

The trained model performed inferences on the validation images, where around 92% of TP (True positives) and 89% of TN (True negatives) were obtained, generating an accuracy of around 91% after training in 100 epochs, with learning rate of 0.001. The program provides the following information to the user: Below 55% indicates a low chance of being a triatomine, 56% to 70% indicates a medium chance, 71% to 85% indicates a moderate chance, 86% to 95% indicates a high chance, and above 96% indicates a very high chance of being a triatomine.  The application of deep learning techniques to automatically classify these vectors speeds up the identification process. Integrated with the Fiocruz Chagas Disease Portal, this system provides specialists with an efficient tool for filtering photos sent by the population via mobile phone, contributing to faster and more accurate decision-making in the diagnosis and control of Chagas disease.  

Palavras Chave

Triatomines; Identification; Deep learning; surveillance; Chagas disease