Comparison of Deep Learning Approaches for Multi-Label Chest X-Ray Classification

Area of Science:

• Medical Imaging
• Computer Science
• Artificial Intelligence

Background:

• Increasing availability of labeled chest X-ray datasets fuels interest in deep learning.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), show promise for medical image analysis.

Purpose of the Study:

• To systematically investigate and compare different deep learning approaches for chest X-ray classification.
• To evaluate the performance of the ResNet-50 architecture and its variations for identifying pathologies in X-ray images.

Main Methods:

• Exploration of transfer learning (with and without fine-tuning) and training from scratch using ResNet-50.
• Inclusion of extended ResNet-50 and a novel architecture integrating non-image data (age, gender, acquisition type).
• Systematic evaluation using 5-fold cross-validation, a multi-label loss function, ROC statistics, and rank correlation analysis.

Main Results:

• A significant variation in performance was observed across different deep learning approaches.
• The X-ray-specific ResNet-38 model, incorporating non-image patient data, demonstrated superior overall performance.
• Class activation maps provided insights into the classification process, highlighting the impact of non-image features.

Conclusions:

• Deep learning models offer powerful tools for chest X-ray classification, but performance varies significantly.
• Integrating non-image data alongside image analysis in specialized architectures like ResNet-38 can substantially improve classification accuracy.
• Further research into explainable AI (XAI) methods like class activation maps is crucial for understanding and trusting AI-driven diagnostic tools.