Deep Learning-Based Networks for Detecting Anomalies in Chest X-Rays.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Chest X-rays are crucial for diagnosing various pathologies like pneumonia and COVID-19.
• Deep learning, especially convolutional neural networks (CNNs), excels at image classification.
• Transfer learning reduces data and computational needs for training deep networks.

Purpose of the Study:

• To investigate deep learning-based neural networks for detecting anomalies in chest X-rays.
• To evaluate different CNN architectures, including transfer learning approaches, on the ChestX-ray14 dataset.
• To compare the performance of pre-trained networks (VGG19, ResNet50, Inceptionv3) with an ad hoc architecture.

Main Methods:

• Implemented transfer learning using pre-trained networks (VGG19, ResNet50, Inceptionv3) with varied classification schemes and data augmentation.
• Developed and evaluated an ad hoc CNN architecture without transfer learning.
• Utilized the ChestX-ray14 database comprising over 100,000 labeled chest X-ray images.

Main Results:

• Transfer learning yielded acceptable results, proving viable for medical image analysis with limited data.
• The ad hoc network demonstrated good generalization capabilities when combined with data augmentation.
• Both approaches showed promise for developing effective chest X-ray pathology detection classifiers.

Conclusions:

• Deep learning, with or without transfer learning, is a promising approach for developing chest X-ray pathology detection classifiers.
• Transfer learning offers a practical starting point for deep network implementation in medical imaging, especially with scarce labeled data.
• The ad hoc network highlights the potential of custom architectures when sufficient data and augmentation are available.