Human-Validated Neural Networks for Precise Amastigote Categorization and Quantification to Accelerate Drug Discovery in Leishmaniasis.

Abstract

Leishmaniases present a significant global health challenge with limited and often inadequate treatment options available. Traditional microscopic methods for detecting Leishmania amastigotes are time-consuming and error-prone, highlighting the need for automated approaches. This study aimed to implement and validate the YOLOv8 deep learning model for real-time detection, quantification, and categorization of Leishmania amastigotes to enhance drug screening assays. YOLOv8 was trained on 470 images from two microscopes, classifying them into categories such as "infected cells," "intracellular amastigotes," "uninfected cells," and "edge cells." The model's performance was compared to human operators using Pearson and Spearman correlation analyses. YOLOv8 achieved strong performance in detecting "infected cells" (AUC = 0.934) and "intracellular amastigotes" (AUC = 0.846). However, challenges remained in differentiating extracellular amastigotes from background noise (AUC = 0.672). Despite these challenges, the YOLOv8 model effectively minimized human variability in drug screening, providing a reliable and efficient tool for the quantification and categorization of Leishmania amastigotes in drug discovery efforts. While further refinements are required to resolve misclassification issues, the model demonstrates significant potential in enhancing both accuracy and throughput in preclinical assays.