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, its reliance on examiners subjective evaluations often leads to inconsistent results. To address this limitation, we propose a deep neural network replicating the agglutination rate estimation in MAT. By leveraging a pre-trained DenseNet121, the network parameters are optimized efficiently during training. We validated our approach using an in-house dataset, and experimental results show that the proposed network achieves accurate agglutination rate estimates. In addition, we employed a standard visualization technique to elucidate the decision-making process, revealing that the network captures image features indicative of leptospire abundance. Overall, our findings suggest that deep learning effectively estimates agglutination rates and that enhancing interpretability supports medical experts in understanding the underlying functionality of deep learning models.