Deep Learning Revolutionizes Parasitic Organism Detection

By Samudrapom DamMar 20 2024Reviewed by Susha Cheriyedath, M.Sc.

In an article recently published in the journal Scientific Reports, researchers investigated the effectiveness of an approach based on deep learning (DL) and fine-tuned optimizer for improving the detection of parasitic organisms in microscopy images.

Limitations of conventional approaches

Parasitic diseases caused by different parasitic organisms like protozoa and ectoparasites pose a significant global health threat, specifically in regions lacking advanced healthcare facilities, affecting millions of individuals. These diseases cause serious and long-lasting health issues, which can also become life-threatening when not treated on time.

Thus, accurate and early parasitic organism detection is crucial for saving lives through effective treatment and timely interventions. Conventionally, parasitic disease diagnosis is time-consuming and laborious as it involves microscopy, molecular techniques, and serological tests. Although these conventional approaches are effective, they also require highly professional and skilled individuals who can properly analyze and understand the disease.

Potential of DL

In recent years, DL and machine learning (ML) models have gained significant attention in the medical sector as they can improve the efficiency and precision of classifying, diagnosing, and detecting several diseases, including parasitic diseases. ML techniques like decision trees, random forests, and support vector machines, and DL techniques like convolutional neural networks and recurrent neural networks can accurately recognize patterns and classify a given data.

These methods have effectively analyzed medical data, including diagnostic images, blood smears, and tissue samples, which makes them suitable for accurately classifying and detecting different parasite diseases. Additionally, ML and DL techniques can be applied to analyze data from emerging technologies like genomic sequencing and rapid diagnostic tests, which make the diagnostic process more accessible and faster.

The proposed DL-based approach

In this study, researchers investigated the role of DL techniques in classifying and detecting several parasitic organisms. A diverse dataset consisting of 34,298 samples of various parasites/parasitic organisms, including Trichomonad, Babesia, Leishmania, Plasmodium, Trypanosome, and Toxoplasma Gondii, and along with host cells like white blood cells and red blood cells, was utilized for this study. The inclusion of white and red blood cells improved the real-world relevance and complexity of the dataset.

Initially, these images were converted from red, green, and blue (RGB) to grayscale, and then morphological features like width, area, height, and perimeter were extracted, facilitating a comprehensive understanding of image characteristics. Eventually, Otsu thresholding and watershed techniques were used to differentiate foreground from background and generate markers on the images for accurately identifying regions of interest.

Deep transfer learning models, including DenseNet169, Xception, MobileNetV2, EfficientNetB0, EfficientNetB3, ResNet152V2, ResNet50V2, InceptionV3, and VGG19, and a hybrid model InceptionResNetV2, were employed. Three optimizers, including Adam, RMSprop, and SGD, were utilized to fine-tune the parameters of these models.

All models were thoroughly evaluated and compared using different performance metrics, including F1 score, precision and recall, root mean square error (RMSE), loss, and accuracy, to determine their effectiveness in parasitic organism classification. The performance of the models was evaluated after applying the three optimizer techniques.

Significance of the study

The work achieved exceptional efficiency and accuracy in categorizing and identifying different parasitic organisms harnessing the capabilities of DL models, coupled with image processing techniques. Specifically, the integration of DL models and the strategic optimization using Adam, SGD, and RMSprop yielded excellent results. The incorporation of these optimizers substantially improved the model performance.

For instance, results demonstrated that EfficientNetB0, InceptionV3, and VGG19 achieved the highest accuracy of 99.91% with 0.09 loss when RMSprop was applied to these models. Similarly, InceptionV3 performs exceptionally well and attained the highest accuracy of 99.91% with 0.98 loss when the SGD optimizer was applied to this model. InceptionResNetV2 displayed the second-best performance by achieving 99.90% accuracy with 1.00 loss.

InceptionResNetV2 attained the highest accuracy of 99.96% with 0.13 loss after the incorporation of the Adam optimizer. InceptionV3 and EfficientNetB3 demonstrated the second- and third-best performance by achieving an accuracy of 99.94% and 99.91%, respectively.

Additionally, the best RMSE and loss values were obtained by InceptionResNetV2, followed by EfficientNetB0 and VGG19. Moreover, the applied models consistently achieved values around 0.99 when they were assessed based on F1 score, recall, and precision. SGD and RMSprop were the fastest optimizers for most of the DL models, with shorter training durations than the Adam optimizer, indicating that SGD and RMSprop were efficient at quickly converging to good model weights.

To summarize, the findings of this study demonstrated the feasibility of using DL models with image processing methods to classify and detect parasitic organisms with high accuracy and efficiency.