Effective residual convolutional neural network for Chagas disease parasite segmentation

Area of Science:

• Medical Imaging
• Computational Biology
• Parasitology

Background:

• Chagas disease, caused by Trypanosoma cruzi, is a neglected tropical disease with significant mortality.
• Early diagnosis is crucial due to asymptomatic infection phases, but current methods are time-consuming and require expertise.
• Automated diagnostic tools are lacking for efficient intervention.

Purpose of the Study:

• To develop an efficient deep learning model for automated semantic segmentation of Trypanosoma cruzi parasites.
• To improve the speed and accuracy of Chagas disease diagnosis in blood samples.

Main Methods:

• An efficient residual convolutional neural network, Res2Unet, was designed incorporating an active contour loss and improved residual connections.
• The model's architecture was inspired by Heun's method for solving ordinary differential equations.
• Trained on 626 blood sample images and validated on 207 images.

Main Results:

• The Res2Unet model achieved a Dice coefficient of 0.84, precision of 0.85, and recall of 0.82.
• Performance surpassed existing state-of-the-art methods for parasite segmentation.
• Demonstrated high accuracy in identifying Trypanosoma cruzi in microscopic images.

Conclusions:

• The developed computational model, Res2Unet, offers a promising solution for prompt Chagas disease diagnosis.
• This AI-driven approach can assist healthcare providers in managing this global health concern.
• Automated parasite detection can significantly expedite the diagnostic process for neglected tropical diseases.