AI revolutionizes malaria diagnosis with 97.57% accuracy using EfficientNet

In a recent study published in Scientific Reports, a team of researchers proposed using an artificial intelligence (AI) tool that uses deep learning to examine red blood cell images in blood smears for the timely detection of malaria.

Study: Efficient deep learning-based approach for malaria detection using red blood cell smears. Image Credit: cones/Shutterstock.com

Background

The World Health Organization report from 2015 shows that in subtropical and tropical regions of the world, the parasite of the genus Plasmodium that causes malaria was responsible for over 400,000 deaths.

Malaria is usually detected through microscopic analysis of blood smear slides, which reveal infected erythrocytes or red blood cells.

Given that regions in Africa, South East Asia, and the Mediterranean experience over 70% of malaria cases, the process of detecting malaria through blood smears becomes very laborious and significantly increases the pathologist’s workload.

AI-based tools involving machine learning and deep-learning approaches have been widely explored in recent studies for automated screening and applications in clinical diagnoses.

However, traditional AI approaches such as neural networks have faced challenges in detecting and identifying malarial parasites in blood smears due to the small size and substantial disparity in blood cells.

Furthermore, these methods still require qualified pathologists for feature vector extraction, making it difficult to automate the screening and detection process completely.

About the study

In the present study, the researchers proposed a deep-learning-based AI tool to detect malaria from images of red blood cells accurately. They also compared the proposed EfficientNet-B2 model against other deep-learning models and used ten-fold cross-validation for efficacy validation.

A dataset consisting of 27,558 blood cell images, of which half were those from uninfected individuals and the other half had parasitized cells, was used in the study. Expert pathologists manually annotated the images.

The preprocessing step involved resizing the images to standardize the size of the images since the model necessitates that the size of the input be fixed or equal.

The images were then split into training and test datasets. The researchers used 80% of the images as the training dataset, while the remaining were used to test the performance and efficacy of the model.

The deep-learning model EfficientNet-B2 used in this study was a Convolutional Neural Networks (CNN) model, which has been widely employed for problems involving image classification.

The model provides accurate classification results by efficiently scaling the images using depth-wise separable convolutions. An added benefit is the small size of the model, requiring lower computing resources.

The researchers used batch normalization to increase the accuracy of the model. This process calculates the mean and standard deviation of each feature using a smaller dataset, which is then used to standardize the input.

A set of classifications for blood cell images obtained from experts was employed to train the deep-learning model to recognize symptoms of malaria.

The study also compared the performance of numerous pre-trained models such as CNN, Visual Geometry Group (VGG16), Inception, DenseNet121, MobileNet, and ResNet, compared to the deep-learning model proposed in this study.

Some of the measures along which the performance of these models was evaluated included false positive, false negative, true positive, and true negative rates, as well as precision, accuracy, and recall.

Results

The study showed that the model proposed in the present study had higher accuracy, area under the curve (AUC), precision, and F1 value, which is the average of precision and recall, compared to the other pre-trained models. Additionally, the testing loss for the proposed model was lower than that of the other models.

After 80% of the dataset was used to train the model, testing the model on the remaining 20% provided an accuracy score of 0.9757, which was higher than the accuracy score obtained when 90% of the dataset was used for the training.

Furthermore, the ten-fold cross-validation indicated that the detection of malaria by the proposed model was highly accurate, with high recall and AUC scores and exceptionally low testing loss.

The model exhibited 98.59% accuracy in detecting cells containing parasites, while the detection of uninfected cells was found to be 100% accurate from the results of the confusion matrix.

Conclusions

Overall, the study showed that the proposed model EfficientNet-B2 exhibited high accuracy and precision in detecting symptoms of malaria from images of blood cells obtained from blood smears. The model outperformed the other existing deep-learning-based models in all the performance parameters.

The researchers believe this model could be employed to improve the accuracy of malaria detection from blood smear samples and significantly reduce the workload of pathologists.