Microscopic parasite malaria classification using best feature selection based on generalized normal distribution optimization

Javeria Amin, Muhammad Almas Anjum, Abraz Ahmad, Muhammad Imran Sharif, Seifedine Kadry, Jungeun KimPeerJ Comput. Sci 2024 · DOI

TL;DR

This paper introduces a novel method for accurately classifying malaria parasites in microscopic images by combining handcrafted and deep learning features. By employing a generalized normal distribution optimization (GNDO) approach for feature selection, the method achieves a remarkable 99% classification accuracy, outperforming existing state-of-the-art techniques and offering a significant advancement in malaria diagnosis.

What Problem Does This Paper Address?

Malaria remains a significant global health threat, and its early and accurate diagnosis is crucial for effective treatment and preventing severe complications. Traditional malaria diagnosis relies on microscopic examination of blood samples, a process heavily dependent on the expertise of pathologists. This manual approach is prone to errors, especially in rural areas with limited access to skilled personnel, and can lead to misdiagnoses and delayed treatment. While image processing and machine learning offer promising avenues for automated diagnosis, challenges persist, including the need for robust feature extraction, effective feature selection, and handling of noisy, low-quality microscopic images. This paper aims to address these challenges by developing a more accurate and efficient automated system for malaria parasite classification.

Methods

The proposed methodology involves a multi-stage approach:

  • Image Preprocessing: A bilateral filter is applied to enhance the quality of microscopic malaria images by smoothing them while preserving edges, thus mitigating noise and improving contrast.
  • Feature Extraction:
    • Handcrafted Features: Pyramid Histograms of Oriented Gradients (PHOG) are extracted, capturing shape and spatial information from the images. These features result in a dimension of N × 300.
  • Deep Features: Features are extracted from the fully connected layers of pre-trained deep learning models, specifically ResNet-18 and ResNet-50. These deep features have a dimension of N × 1,000 each.
  • Feature Fusion: The handcrafted (PHOG) and deep features (from ResNet-18 and ResNet-50) are serially fused to create a comprehensive feature vector with a dimension of N × 2,300.
  • Feature Selection: The Generalized Normal Distribution Optimization (GNDO) algorithm is employed to select the most informative features from the fused set. This optimization process reduces the feature dimension to N × 498. GNDO is based on the normal distribution theory and iteratively searches for the optimal feature subset.
  • Classification: The selected features are then used to classify malaria cells using Support Vector Machine (SVM) and neural network classifiers.

Datasets

The study utilizes a publicly available RGB microscopic images parasite malaria dataset from the National Library of Medicine. This dataset contains two classes: infected and uninfected. Each class comprises 13,779 images, with dimensions of 256 × 256 × 3 pixels. The dataset is divided into ten folds for 10-fold cross-validation during experiments.

Key Insights and Ideas

  • Synergistic Feature Fusion: Combining handcrafted features (PHOG) with deep features from pre-trained CNNs (ResNet-18, ResNet-50) proves more effective than using either type of feature alone. This highlights the complementary nature of different feature extraction approaches for malaria image analysis.
  • GNDO for Optimal Feature Selection: The application of Generalized Normal Distribution Optimization (GNDO) for feature selection is a key innovation. It effectively prunes the high-dimensional fused feature space to a smaller, highly informative subset, leading to improved classification performance and reduced computational complexity.
  • Bilateral Filtering's Impact: The initial application of a bilateral filter significantly improves image quality, which is crucial for subsequent feature extraction and ultimately contributes to higher classification accuracy, especially given the inherent noise in microscopic images.
  • High Accuracy Achieved: The method consistently achieves very high classification accuracies, demonstrating its robustness and potential for real-world diagnostic applications.

Key Results

The proposed method achieved a classification accuracy of 99%. This is a significant improvement over previously published works. Specifically:

  • Individual feature extraction methods yielded accuracies like 0.89 for PHOG, 0.87 for ResNet-18, and 0.82 for ResNet-50 when using linear SVM.
  • Fusion of ResNet-18 and ResNet-50 features alone achieved an accuracy of 0.90.
  • After feature selection using GNDO, the best features were used for classification, leading to the reported 99% accuracy.
  • In terms of ROC analysis, the wide neural network classifier achieved a mean ROC of 0.98 ± 0.00, and the SVM quadratic classifier achieved 0.97 ± 0.01.

Limitations and Future Work

  • Limitations: The current work is limited to binary classification (infected vs. uninfected). The authors acknowledge that real-world patient data from various acquisition sources might present additional complexities.
  • Future Work: The researchers suggest extending this research to multiclassification of malaria parasites using real patient data. Furthermore, they propose exploring quantum machine learning for malaria classification, given the growing importance of quantum computing.

Summary updated

BibTeX
@article{amin2024microscopic,
  author = {Javeria Amin and Muhammad Almas Anjum and Abraz Ahmad and Muhammad Imran Sharif and Seifedine Kadry and Jungeun Kim},
  title = {{Microscopic parasite malaria classification using best feature selection based on generalized normal distribution optimization}},
  journal = {PeerJ Comput. Sci.},
  year = {2024},
  doi = {10.7717/peerj-cs.1744},
}
Figure