Feasibility of a convolutional neural network (CNN)-based approach for the detection of Leishmania parasites in microscopic fields as an adjunct diagnostic tool for leishmaniasis

Abstract

Leishmaniasis, caused by protozoan parasites of the genus Leishmania, is ranked among the top 10 neglected tropical diseases (NTDs) by the World Health Organization and constitutes a significant global public health burden. In Sri Lanka, cutaneous leishmaniasis is the predominant clinical form, and the disease is considered endemic, yet the absence of a national surveillance program and the scarcity of trained microscopists remain critical barriers to timely and accurate diagnosis. Conventional microscopy, while remaining the primary diagnostic standard, is inherently operator-dependent and susceptible to misinterpretation by less experienced personnel. The present study aimed to evaluate the feasibility of a convolutional neural network (CNN)-based deep learning model, built on the TensorFlow framework, for the automated detection of Leishmania amastigotes in stained microscopic field images. A dataset comprising 161 microscopic field images, including 91 positive images for Leishmania amastigotes and 70 negative images, was used to evaluate model performance. The model correctly classified 88 of 91 positive images (sensitivity: 96.7%) and 65 of 70 negative images (specificity: 92.9%), achieving an overall diagnostic accuracy of 95.0%. The positive predictive value and negative predictive value were 94.6% and 95.6%, respectively. Receiver operating characteristic (ROC) curve analysis demonstrated excellent discriminatory performance, with an area under the curve (AUC) of 0.999. Notably, the model demonstrated the capacity to discriminate Leishmania amastigotes from morphologically similar artefacts, including platelet clumps, a recognized source of false-positive diagnoses in routine microscopy. Performance benchmarks were comparable to, and in some instances exceeded, those reported in previous deep learning studies for Leishmania parasite detection and other haemoparasite classification tasks. These findings provide compelling evidence for the feasibility of CNN-based automated image analysis as a cost-effective, and accurate adjunct or alternative to conventional microscopy for leishmaniasis diagnosis in resource-limited settings. Validation using larger, geographically diverse datasets and prospective clinical workflow evaluation is recommended as the next step toward clinical implementation.