Development of an AI model for pneumothorax imaging: Dataset and model optimization strategies for real-world...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Deep Learning for Diagnostics

Background:

• Traditional pneumothorax diagnosis relies on subjective interpretation of chest X-rays, which can be impacted by radiologist fatigue and experience.
• Developing automated systems can enhance diagnostic efficiency, accuracy, and reduce workload.

Purpose of the Study:

• To develop and evaluate an AI-assisted system for diagnosing pneumothorax using deep learning and chest X-ray images.
• To improve the accuracy and efficiency of pneumothorax detection compared to traditional methods.

Main Methods:

• Utilized the DenseNet121 deep learning model trained on 6888 chest X-ray images from a Taiwanese medical center.
• Employed image preprocessing techniques including normalization and data augmentation.
• Trained the model using stochastic gradient descent and evaluated performance using accuracy, sensitivity, specificity, and AUROC.

Main Results:

• Initial testing showed high AUROC values (94.52-97.21%) for specific groups.
• Application to a larger clinical dataset resulted in a significant drop in AUROC (62.55%) with numerous false positives.
• Retraining the model with 1000 false positive images improved the AUROC to 85.53%.

Conclusions:

• The AI model demonstrates potential for pneumothorax detection but is sensitive to data diversity, image quality, and clinical complexity.
• Further improvements may require attention mechanisms or region proposal networks to handle complex cases.
• Expanding datasets and optimizing preprocessing are crucial for enhancing clinical performance.