Attacking the out-of-domain problem of a parasite egg detection in-the-wild

Area of Science:

• Computer Science, Machine Learning
• Parasitology, Diagnostics

Background:

• Machine learning models struggle with out-of-domain (OO-Do) data, where test data differs from training data, limiting real-world deployment.
• Accurate parasite-egg detection is crucial for public health, but existing models face challenges with diverse, real-world conditions.

Purpose of the Study:

• To address the OO-Do problem in parasite-egg detection using object detection.
• To introduce a novel dataset and a data-driven framework for robust parasite-egg recognition in real-world applications.

Main Methods:

• Development of the 'In-the-wild parasite-egg' dataset (1,552 images) collected via chatbot from 222 students.
• Proposal of a data-driven framework using publicly available datasets to train models on both in-domain and OO-Do concepts.
• Evaluation of integration strategies and thresholding for a two-step parasite-egg detection approach.

Main Results:

• The proposed framework, integrating an OO-Do-aware classification model post-object detection, significantly improved parasite-egg detection.
• Achieved 7.37% and 4.09% F1-score improvements over baselines on the Chula+Wild and In-the-wild parasite-egg datasets, respectively.
• Demonstrated enhanced model robustness to OO-Do data through optimized thresholding strategies.

Conclusions:

• The developed framework effectively mitigates the OO-Do problem in parasite-egg detection.
• The 'In-the-wild parasite-egg' dataset provides a valuable resource for evaluating OO-Do-aware models in realistic settings.
• The findings support the practical application of machine learning in parasitology diagnostics.