Deep learning-based classification for lung opacities in chest x-ray radiographs through batch control and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Classifying lung opacities from chest x-rays is crucial for diagnosing pneumonia.
• Existing datasets often suffer from class imbalance, posing challenges for model training.
• The Radiological Society of America (RSNA) 2018 Pneumonia Detection Challenge provided relevant datasets.

Purpose of the Study:

• To implement a system for classifying lung opacities from frontal chest x-ray radiographs.
• To propose and evaluate a novel training method to address class imbalance in medical imaging datasets.
• To enhance the performance and adaptive sensitivity of deep-learning models for pneumonia detection.

Main Methods:

• Utilized convolutional neural networks (CNNs) for image classification.
• Developed a training procedure named 'batch control' to manipulate data distribution within training batches.
• Applied the batch control method to the RSNA 2018 Pneumonia Detection Challenge dataset.

Main Results:

• The implemented CNN system demonstrated practical utility in classifying lung opacities.
• The batch control method effectively regulated and stabilized deep-learning model performance.
• The batch control method proved advantageous for sensitivity regulation in class-unbalanced datasets.

Conclusions:

• Convolutional neural networks are effective for lung opacity classification on chest x-rays.
• The batch control training method offers significant benefits for optimizing models with imbalanced datasets.
• This approach enhances model sensitivity and adaptability for specific clinical applications.