Pneumonia and pneumothorax detection: A multi-factor evaluation of chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Chest radiography is crucial for diagnosing lung conditions like pneumonia and pneumothorax.
• Existing diagnostic systems face challenges with class imbalance and limited supervision in large datasets.
• Vision Transformers (ViT) offer a promising deep learning approach for image analysis tasks.

Purpose of the Study:

• To develop and evaluate a Vision Transformer (ViT) based diagnostic system for pneumonia and pneumothorax detection.
• To address challenges of multi-label imbalance and optimize model performance using specific training strategies.
• To enhance interpretability of diagnostic predictions through attention-based visualization.

Main Methods:

• Utilized the NIH ChestX-ray14 dataset for training and evaluation.
• Implemented radiograph-specific augmentation and WeightedRandomSampler for handling class imbalance.
• Employed CosineAnnealingWarmRestarts scheduling and optimized classification heads and disease-specific thresholds.
• Evaluated model performance using AUC, sensitivity, specificity, precision, and F1-score.

Main Results:

• ViT models achieved 70-75% accuracy and 0.63-0.71 AUC for pneumonia and pneumothorax detection.
• Attention-based visualization improved interpretability by highlighting key radiological areas.
• The system demonstrated enhanced detection of rare conditions.

Conclusions:

• ViT performance is significantly improved by medical-focused data preparation and training.
• The developed system shows potential for assisting radiologists in high-volume, resource-constrained settings.
• ViT-based systems offer a viable tool for improving diagnostic accuracy and interpretability in medical imaging.