A Patch-based Microscopic Image Analysis for Visceral Leishmaniasis Screening Using a Deep Metric Learning Approach

Abstract

Human Visceral Leishmaniasis (VL) is a fatal disease in over 95% of untreated cases and predominantly affects populations with limited access to healthcare. Parasitological techniques are the gold standard for diagnosing VL. It involves the direct microscopic examination of the parasite amastigotes approximately 2–4µ m in diameter. However, this process can be time-consuming and labor-intensive, necessitating a high level of expertise. We propose a novel approach to the detection of these amastigotes by combining deep metric learning with supervised classification techniques. We outperform the state-of-art for this detection problem achieving an f1-score of approximately 99% by tackling poor segmentation and class imbalance drawbacks.