Deep Convolutional Neural Network applied to Chagas Disease Parasitemia Assessment

Title: Deep Convolutional Neural Network applied to Chagas Disease Parasitemia Assessment

Authors: André Pereira, Alexandre Pyrrho, Daniel Vanzan, Leonardo Mazza, José Gabriel Gomes

Abstract: Chagas Disease is a tropical parasitic disease endemic to Latin America, and it is caused by ​ Trypanosoma cruzi. It occurs in two phases. The acute phase takes place shortly after infection, and it is characterized by fever, lymphadenopathy, and chagoma symptoms. The chronic phase, which happens from a few months up to several years after infection, is generally asymptomatic, but it may also be associated with megacolon, megaesophagus, or cardiomegaly symptoms. Other heart illness symptoms may be present as well. In the acute phase, standard diagnosis is based on T. cruzi visualization through microscopy applied to blood smear slides. In the present work, we apply a deep convolutional neural network (namely, a pre-trained Mobile NetV2 feature extractor followed by a fine-tuned single-neuron top classifier) to the binary classification of image tiles of size 224 × 224 × 3, which are extracted from acute-phase blood smear samples. The data set corresponds to blood smear sample images taken from twelve different slides. We achieve 96.4% accuracy on a balanced validation subset within the twelve-slide data set. The respective precision, sensitivity, and F1-score values are 95.4%, 97.6%, and 96.5%. In a cross-validation experiment with five folds inside the twelve-slide data set, validation accuracy varies from 88% to 98%. From image tiles extracted from a thirteenth blood smear slide (i.e. tiles outside the train/validation sets), we estimate test accuracy equal to 72.0%, which suggests that data set size and overtraining issues must be addressed in future work.

Key-words: Chagas disease, ​Trypanosoma ​cruzi​, blood smear samples, deep convolutional neural networks

Pages: 8

DOI code: 10.21528/CBIC2019-119

PDF file: CBIC2019-119.pdf

BibTeX file: CBIC2019-119.bib