Emulating visual evaluations in the microscopic agglutination test with deep learning

Abstract

The Microscopic Agglutination Test (MAT) is widely recognized as the gold standard for diagnosing zoonosis leptospirosis. However, the MAT relies on subjective evaluations by human experts, which can lead to inconsistencies and inter-observer variability. In this study, we aimed to emulate expert assessments using deep learning and convert them into reproducible numerical outputs to achieve greater objectivity. By leveraging a pre-trained DenseNet121, the network benefits from better initialization, facilitating more effective training. We validated our approach using an in-house dataset, and the experimental results demonstrate that the proposed network achieved accurate agglutination rate estimates. In addition, we employed UMAP, a dimensionality reduction technique, to visualize the learned feature representations, revealing that the network captured image features indicative of Leptospira abundance. Overall, our findings suggest that deep learning can consistently estimate agglutination rates in a manner that approximates expert evaluations and that enhancing interpretability provides visual cues that could aid in understanding the behavior of deep learning models, potentially facilitating future clinical integration.