Development of OO-Do-Aware Parasite Egg Detection

Abstract

The out-of-domain (OO-Do) problem occurs when a machine learning model is presented with test data that does not belong to any of the classes present in the training data. As a result, the model will always render an incorrect prediction, predict OO-Do as one of the trained classes. Parasitic infections can be a significant public health issue, and detecting and identifying parasite eggs in images can be a helpful technology for early diagnosis and treatment. However, a parasite egg detection model may face challenges when presented with out-of-domain (OO-Do) data, which includes images of unrelated objects such as cats, trees, or other irrelevant content. Previous research developed techniques for detecting out-of-domain samples in object detection to overcome this issue. These methods typically involve modifying a pre-trained model for object detection to improve its ability to detect samples outside the domain it was trained. To retain the performance of the original model while improving its ability to detect objects in OO-Do samples, we adopted a two-step approach to address the challenge of the out-of-domain samples. The first step involves classifying the test images using threshold strategies. The second step is employing object detection techniques to detect further and verify the out-of-domain samples. Object detection without threshold strategy and a two-step approach using SoftMax threshold achieved an F1-score of 77.30% and 70.60%, respectively. For out-of-domain image awareness, a two-step approach using SoftMax threshold obtained 57.97% F1-score compared to 29.94% F1-score of object detection without threshold strategy. This suggests that the proposed approach effectively addressed the out-of-domain problem in the context of parasite egg detection.