Enhancing leptospirosis screening using a deep convolutional neural network with microscopic agglutination test images.

Abstract

Leptospirosis poses substantial challenges to global public health. In Malaysia, leptospirosis is endemic, with annual cases peaking during the monsoon season. The microscopic agglutination test (MAT) is the gold-standard serological method for confirmation of leptospirosis. However, it is labor-intensive and time-consuming, as it relies on the subjective interpretation of medical lab technicians. This study describes the development of a semiautomated workflow for Leptospira screening by integrating a TensorFlow and custom-designed Keras-based Deep Convolutional Neural Network (DCNN) with conventional MAT. We used a dataset of 442 positive and 442 negative MAT images, which consisted of a mixture of Leptospira serovars from Malaysia to train the model. The model was subjected to hyperparameter tuning, which modulated the number of convolutional layers, filters, kernel sizes, units in dense layers, activation functions, and learning rate. Verification of our tested model compared to the verified patient MAT results achieved the following metrics: a Precision score of 0.8125, a Recall of 0.9286, and an F1-Score of 0.8667. Combining our model with the current Malaysia Leptospira workflow can significantly speed up, reduce inaccuracies, and improve the management of leptospirosis. Furthermore, the application of this model is practical and adaptable, making it suitable for other labs that observe MAT as their Leptospira diagnosis. To our knowledge, this approach is Malaysia's first hybrid diagnostic approach for Leptospira diagnosis. Scaling up the dataset would enhance the model's accuracy, making it adaptable in other regions where leptospirosis is endemic.