Deep Transfer Learning for the Multilabel Classification of Chest X-ray Images.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiology informatics

Background:

• Chest X-ray (CXR) is a primary diagnostic tool for thoracic conditions.
• Deep convolutional neural networks (CNNs) show promise for automated CXR interpretation.
• Limited annotated CXR datasets pose a challenge for training robust deep learning models.

Purpose of the Study:

• To optimize deep learning models for chest X-ray disease identification using transfer learning.
• To evaluate the impact of various transfer learning strategies on CXR classification performance.
• To investigate the influence of different source datasets and CNN architectures.

Main Methods:

• Utilized a private dataset of 1630 CXR images with disease labels.
• Employed deep CNN models for feature extraction and disease identification.
• Implemented transfer learning by reusing pretrained weights from source datasets (ImageNet, ChestX-ray, CheXpert) via finetuning and layer transfer.
• Assessed different integration methods (initiating, concatenating, co-training) and backbone models (ResNet50, DenseNet121).

Main Results:

• Transfer learning with model finetuning generally yielded superior prediction models.
• Combining ImageNet with CheXpert outperformed using CheXpert alone.
• ResNet50 excelled in initiating transfer learning, while DenseNet121 was better for concatenating and co-training.
• Transfer learning using multiple source datasets proved more effective than using a single source dataset.

Conclusions:

• Transfer learning significantly enhances prediction capabilities for chest X-ray analysis.
• Model finetuning and the use of multiple diverse source datasets are recommended for optimal performance.
• Transfer learning offers a viable approach to reduce computational costs and improve diagnostic accuracy in CXR interpretation.