Performance validation of deep-learning-based approach in stool examination

Area of Science:

• Computational biology and bioinformatics
• Medical diagnostics and parasitology

Background:

• Human intestinal parasitic infections (IPI) affect billions globally, causing significant mortality.
• Traditional diagnostic methods like Kato-Katz and FECT are standard but have limitations.
• This study investigates deep learning for improved intestinal parasite identification.

Purpose of the Study:

• To evaluate the performance of deep learning models for identifying intestinal parasites.
• To compare the diagnostic accuracy of deep learning approaches against human experts.

Main Methods:

• Human experts utilized FECT and MIF techniques for ground truth.
• Images from modified direct smears were used to train (80%) and test (20%) deep learning models (YOLOv4-tiny, YOLOv7-tiny, YOLOv8-m, ResNet-50, DINOv2).
• Performance was assessed using confusion matrices, ROC, PR curves, Cohen's Kappa, and Bland-Altman analyses.

Main Results:

• Deep learning models, particularly DINOv2-large and YOLOv8-m, demonstrated high accuracy (up to 98.93%) and specificity (up to 99.57%).
• Models showed strong performance in identifying helminthic eggs and larvae due to distinct morphology.
• All models achieved high agreement (Kappa > 0.90) with medical technologists; Bland-Altman analysis indicated good agreement and minimal bias.

Conclusions:

• Deep learning approaches show significant potential for automated parasite identification, enhancing diagnostic accuracy for IPI.
• Integrating deep learning can lead to earlier detection and more effective interventions for parasitic infections.
• This technology represents a substantial advancement in improving IPI diagnostic procedures.