Dynamic learning for imbalanced data in learning chest X-ray and CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models require large annotated datasets, which are often limited and unbalanced for novel research areas like viral epidemics.
• Detecting lung diseases from chest X-rays and CT scans is challenging due to data scarcity and class imbalance, particularly for rare or severe cases.

Purpose of the Study:

• To develop a robust deep learning framework for accurate lung disease detection from medical images, specifically addressing the challenge of imbalanced datasets.
• To enhance the identification of minority classes (e.g., significant instances of novel illness) within limited medical image datasets.

Main Methods:

• A novel class balancing algorithm, 3-Phase Dynamic Learning (3PDL), was developed to handle imbalanced datasets.
• A parallel Convolutional Neural Network (CNN) model, Hybrid Feature Fusion (HFF), was employed for multi-modal image analysis (X-ray and CT).
• Support Vector Machine (SVM) was utilized for image categorization within the clustering process, and probabilistic modeling represented data characteristics.

Main Results:

• The proposed 3PDL and HFF models achieved a high F1 score of 96.83% and a precision of 96.87%.
• The system demonstrated outstanding accuracy and generalization capabilities in identifying lung disease signs.
• The imbalance-based sample analyzer effectively identified minority categories, improving classification performance.

Conclusions:

• The developed technique successfully addresses class imbalance in medical image analysis for lung disease detection.
• The high performance of the 3PDL and HFF models suggests their utility as a supportive tool for physicians and pathologists.
• This approach offers a promising solution for improving diagnostic accuracy in scenarios with limited and unbalanced medical data.