Automatic Detection of Trypanosomosis in Thick Blood Smears Using Image Pre-processing and Deep Learning

Taewoo Jung, Esla Timothy Anzaku, Utku Özbulak, Stefan Magez, Arnout Van Messem, Wesley De Neve

Abstract: Trypanosomosis, which is caused by the Trypanosoma parasite, is an infectious disease that affects both humans and animals. Today, the microscopic examination of a Giemsa or Wright stained blood smear from an infected individual is the standard procedure for diagnosis because of the straightforward nature of sample preparation. Unfortunately, this method is labor-intensive and prone to error, particularly resulting in false-negative scoring when parasite levels are low during chronic infections. Automating the detection of parasites in blood smear images can overcome sensitivity limitations related to a microscopic examination. We therefore propose a deep learning approach that aims at automatically classifying microscope images in terms of parasite presence or absence. To that end, we applied a ResNet18 model using a pre-processed dataset derived from microscope videos of unstained thick blood smears, with the blood smears originating from a mouse infected with Trypanosoma brucei. Our pre-processing strategy mainly involved image cropping and the application of a thresholding algorithm for facilitating effective model training. Moreover, our thresholding approach made it possible to observe a positive correlation between the percentage of parasite-related pixels in an image and the classification effectiveness.