Enhancing Detection of Leishmania spp. Amastigotes in Canine Lymph Node Smear Images: Evaluating the Effectiveness of Synthetic Data in Augmenting Existing Datasets

Dimitrios Tsikos, Irene Chatzipanagiotidou, Dimitris Dimitriadis, Constantina N. Tsokana, George Valiakos, Labrini V. Athanasiou, Grigorios Tsoumakas

Abstract: Leishmaniosis is a parasitic mammalian disease that severely affects humans and dogs. Early diagnosis is crucial and associated with improved prognosis and treatment outcomes. A key diagnostic component is the detection of Leishmania amastigotes, the etiological agent of the disease, in cytologic preparations via microscopy. However, reliance on operator expertise limits its accesibility in veterinary clinics. Deep learning offers a promising approach for automating Leishmania amastigote detection, yet data limitations and the time-consuming, error-prone nature of real data annotation process remain significant challenges. This study explores the use of synthetic data to address these challenges and improve deep learning performance in detecting Leishmania amastigotes in microscopic images from canine lymph node aspirates. We propose an automated, two-stage synthetic data generation approach. First, structured representations of healthy and infected cells are created based on real microscopy data, incorporating randomized morphological features and material properties to mimic optical characteristics. Then, these elements are assembled into composite images with controlled variations in spatial arrangement, lighting, and perspective to enhance dataset diversity. The final output is annotated images designed for training object detection models. By supplementing real datasets with synthetic images, we address data scarcity and imbalance issues, improving model accuracy and generalization. Our results show that incorporating synthetic data significantly enhances deep learning models’ ability to detect Leishmania amastigotes, offering a promising solution for veterinary diagnostics. Additionally, we introduce a new dataset that combines both original and synthetic data, contributing to further research into this important zoonotic disease.

External IDs:dblp:conf/pkdd/TsikosCDTVAT25