Automatic detection of the parasite Trypanosoma cruzi in blood smears using a machine learning approach applied to mobile phone images

Samples analyzed

A total of 33 slides with thin blood smears of Swiss mice experimentally infected with T. cruzi Y strain at acute phase of infection were prepared for image annotation and analysis. Sample preparations were obtained from animals of the Laboratory of Chagas disease, Federal University of Ouro Preto where the T. cruzi strain was maintained through successive blood passages in mice. The Ethics Committee for the Use of Animals (CEUA) at the Federal University of Ouro Preto, Laboratory of Chagas Disease, Minas Gerais, Brazil provided approval for this research (CEUA No. 2015/50).

Object segmentation

We standardized the resolution of the images to 768 × 1,024 pixels² before segmenting the parasites. We applied a graph-based segmentation method (Felzenszwalb & Huttenlocher, 2004). In this process, a graph-based representation of the image is defined were pixels corresponds to vertices and neighboring pixels are connected edges. The contours of the regions of interest with the parasites are obtained by selecting the edges between the different regions based on the differences in intensity between the regions, and the difference in intensity between the pixels within each region. As a result, we observed the whole parasite cell body segmented within each region.

Next, we cropped a 100 × 100 pixels² region around each parasite based on manual position annotation. Only the regions of interest with the parasite were selected for processing and feature extraction. This procedure resulted in 1,314 parasites. We selected other segments from the images with features very similar to the parasite using nearest neighbors technique. In other words, we selected objects that do not belong to parasites and clustered under the label “Unknown”. In this way, we were able to obtain the same number of objects (T. cruzi or unknown) for the object classification task.

Feature extraction

After the segmentation of the parasites, we performed the conversion from the RGB color space to CIEL*a*b* color space for object’s feature extraction (see the Supplemental Information file for details). These features are object descriptors and are classified as geometric (perimeter, area, circularity, thickness ratio, centroid to contour distances, major and minor axis, aspect ratio, etc.); curvature (entropy, bending energy, standard deviation, and variance) (Costa & Cesar Jr, 2009); texture (color co-occurrence matrices, entropy, inverse difference moment, angular second moment, contrast, correlation) color (mean, median, mode, amplitude, and variance) (Gui et al., 2013; Palm, 2004) descriptors and Hu’s invariant moments (Huang & Leng , 2010).

Feature selection

To increase performance, reduce the noise, and avoid overfitting, we applied either Principal Component Analysis (PCA). In this manner, we reduced the number of features used by the algorithm during the training. The dimension reduction with PCA was performed by keeping the 16 highest eigenvalues of the covariance matrix derived from the feature space data. The proportion of variance of the 16 principal components correspond to 95% of the original variance. Therefore, we obtained a new feature space data that consisted of eigenvectors of the principal components. We obtained the final dataset by multiplying the transposed matrix of eigenvectors by the input data matrix centered by the mean.

Object classification

We applied supervised learning classification with support vector machines (SVM), k-Nearest Neighbours (KNN) and Random Forest (RF) as they presents good generalization with a small dataset (Chen et al., 2020; Hsu, Chang & Lin, 2016; Ben-Hur & Weston, 2010; Cunningham & Delany, 2007; Breiman, 2001).

In the SVM method, each sample is represented by a point in an n-dimensional space, where n is the number of object features and its values are the coordinates of the point, including its class. The classifier searches for the optimal hyperplane that will be used to split the points belonging to distinct classes. In the KNN method, a sample is classified to the data label which has the most representatives within the k nearest neighbors features of the sample. In the RF method, the classification of each object happens through a combination of decision trees. The classification performance of each tree in the ensemble is improved by bootstrapping sampling and aggregation from the training set. The final classification is made by averaging each decision tree probabilistic prediction. In addition, we also performed classification task using an ensemble of these methods based on “soft” voting classifier. The voting classifier combines the results of the different predictors and use the average of predicted probabilities to assign the class labels to each sample.

We used Python’s scikit-learn library to train and validate the models (Pedregosa et al., 2011). We performed object feature data standardization before classification since data presented different orders of magnitude. We assessed the classification performance of the methods by comparing the samples in the feature space for a given class with all other samples. After finding the highest accuracy value for each model, we assessed the models in the validation set.

Statistical analysis

The performance of the constructed models was assessed by Receiver Operating Characteristic (ROC) curve analysis, where the average of the area under the curve (AUC) was calculated for the quantification in both training set and validation sets. We also evaluated the model performance based on sensitivity, specificity, precision, and F1-score.

Image analysis pipeline for detecting T. cruzi

We developed and tested an image analysis pipeline which was based on a ML model to detect T. cruzi in the blood during acute infection (Fig. 1). Initially, the images were prepared from blood samples collected from female Swiss mice acutely infected with the Y strain of Trypanosoma cruzi. The collected blood sample smears were stained by the Giemsa method (Vallada, 1999). This technique allows the observation of parasites with oil immersion objectives. Regions of nucleic acid-rich in adenine-thymine bond tend to get darker. On the other hand, regions rich in cytosine-guanine bonds are less prone to embody the Giemsa stain and tend to present clearer stains.

A total of 33 slides with thin blood smears from different animals were stained with Giemsa. About 20 fields of view that correspond to each of the analyzed images were captured from each slide. We manually acquired 674 images from the slides under 100x objective (CFI E Plan Achromat 100x Oil, 1.25 NA/0.23 W.D.) in an optical microscope (Nikon Eclipse E200) with a cell phone camera (Morola Moto G4) attached to the eyepiece (CFI E 10x, F.N. 20 mm) (Fig. 2). The camera was configured with the macro focus and other configurations were set to automatic for acquisition. With these settings, the images were acquired with a resolution of 3,456 × 4,608 pixels², a field of view with a diameter of 0.2 mm resulted in a pixel lateral size of approximately 0.06 µm. A few images were acquired at 2,448 × 3,264 pixels² due to change in camera configuration. This was useful for classifier test at different input image resolutions. Images were stored in JPEG format, at 100% quality, and file names standardized according to unique identifiers.

Parasites were microscopically identified by two specialist researchers in T. cruzi. A total of 1,314 parasites were observed. We then marked the position of each nucleus or cell body of the T. cruzi found in the images. The position of the objects of interest (parasite) in the image was obtained by a point-and-click event using an “in-house software”. The image identifier, the pointed object, and the coordinates in X and Y axis information were extracted and stored in a database (Data S1).

To detect the objects of interest, we applied a graph-based approach (Felzenszwalb & Huttenlocher, 2004). Segmented regions with more than 3,000 pixels were considered not to contain the objects of interest. After image segmentation, objects were cropped in a 100 × 100 pixel² sub-region around its X and Y coordinates and labeled into two classes: parasite and unknown (Fig. 3). The segmented regions containing the the T. cruzi trypomastigote form were labeled as “parasite” (Fig. 4A). The segmented regions that do not contain a parasite or that are over-segmented were labeled as “unknown” (Fig. 4B).

We selected a set of regions labeled as “unknown” to train and validate the classifier method. The regions were selected based on the features values closest to the regions labeled as “parasite” using the nearest neighbors method. In this way we achieved a total of 1,314 segments marked as parasite and the same number of segments marked as unknown (Table 1).

To identify the parasite in the image sub-region, we first extracted features from these regions. These features represent a description of the object’s morphology (geometry and curvature), as well as color and texture. We also calculated Hu’s invariant moments to capture information regarding shape and intensity regardless of the object’s position and size (Table 2). In total, we extracted 49 features from the segmented objects of each class (Data S2).

We then split the region’s feature data into two sets based on the number of acquired images. This resulted in a proportion of about 80% of regions for the training set (n = 2181) and 20% for test (n = 447). We also applied principal component analysis (PCA) to reduce the number of features and test whether it may improve the classification performance. The proportion of variance of the 16 principal components corresponds to 95% of the original variance of the data (Fig. S1). Therefore, we used 16 features of the transformed values matrix to train the model.

Object classification task presented acceptable performance

We developed a classifier algorithm based on the object features. We observed better performance in the classification task without the feature selection. The models trained on features data selected by PCA were not able to generalize well on the test set, indicating overfitting (Table S1). Without the feature selection, the random forest classifier model presented acceptable performance with an accuracy of 99.7% and area under the ROC curve of of 1.0 in the training set, all the while presenting accuracy of 89.5% and AUC of 0.942 in the test set (Table 3). The voting classifier presented accuracy of 93.3% and AUC of 0.978 in the training set, and accuracy of 79.4% and AUC of 0.884 in the test set (Fig. 5).

The voting classifier confusion matrix presented sensitivity and specificity values of 76.8% and 81.8%, respectively (Table 4). The lower performance presented by the voting classifier is because the ensemble’s prediction is based on the average of the prediction probabilities of each classifier. Since the SVM and KNN classifiers presented lower performance individually (both presented sensitivity of 69.7% and specificity of 75.4%, Table 3) the prediction of the ensemble was lower. On the other hand, the RF classifier had a sensitivity of 90.5% and a specificity of 88.6% (Table 5).

The objects found in the images were then marked by the algorithm as parasite. We found two problems in the parasite recognition task. The first is related to the high rate of false positives (Fig. 6A). The regions of the images with leukocytes or high density of red cells showed over-stained areas. These areas were difficult to classify by the algorithm. The second problem was the false negative rate (Fig. 6B). Parasites in regions of the image that presented low contrast or low sharpness (“out of focus”), most commonly found at the edges of the field of view, were not recognized by the algorithm. This second problem is much more significant, since undiagnosed Chagas disease can put a person’s life at risk.