Machine Learning Model for Chest Radiographs: Using Local Data to Enhance Performance.

Area of Science:

• Artificial Intelligence in Radiology
• Machine Learning for Medical Imaging
• Radiographic Interpretation

Background:

• Generalizability of machine learning models across institutions remains a challenge for widespread implementation.
• Existing AI models for chest radiograph screening often require adaptation to local data characteristics.
• Developing robust AI tools for medical diagnostics necessitates addressing performance variations between different healthcare settings.

Purpose of the Study:

• To develop and evaluate a machine learning model for screening chest radiographs across 14 diagnostic labels.
• To assess the impact of fine-tuning an existing model with local institutional data on its performance.
• To test the hypothesis that local data fine-tuning improves model generalizability and accuracy.

Main Methods:

• An ensemble of neural networks was trained on open-source chest radiograph datasets for 14-label detection.
• The model underwent fine-tuning using 4,510 local radiograph studies, with radiologist reports serving as the gold standard.
• Performance was evaluated on 802 local radiographs using receiver-operator characteristic curves and statistical analysis.

Main Results:

• The fine-tuned model achieved an area under the curve greater than 0.75 for 12 out of 14 pathology labels.
• Statistically significant improvements in overall model performance were observed after fine-tuning (P < .01).
• The model demonstrated enhanced detection capabilities for six specific pathology labels post-fine-tuning.

Conclusions:

• A machine learning model for simultaneous detection of 14 chest radiograph labels was successfully developed and refined.
• Fine-tuning with local data demonstrably improved the model's accuracy and diagnostic performance.
• This strategy offers a practical method to enhance the generalizability and local utility of AI models in diverse clinical environments.