Pulmonary abnormality screening on chest x-rays from different machine specifications: a generalized AI-based image...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning for Medical Diagnosis

Background:

• Chest x-rays are crucial for pulmonary abnormality screening.
• AI models trained on specific x-ray equipment often fail to generalize to images from different machines.
• Variations in x-ray machine specifications significantly impact image characteristics.

Purpose of the Study:

• To develop and evaluate a novel image manipulation pipeline to improve the generalization of AI models for chest x-ray analysis.
• To address the challenge of AI model performance degradation due to diverse x-ray machine specifications.
• To enhance the clinical applicability of AI-driven pulmonary abnormality screening.

Main Methods:

• Collected 15,010 chest x-rays from five institutions, utilizing systems with different generators/detectors.
• Developed an AI model for pulmonary abnormality classification using x-rays from a single system.
• Compared the proposed XM-pipeline with conventional methods (HE, CLAHE, UM) using multicenter external datasets.

Main Results:

• The XM-pipeline demonstrated superior performance across datasets from various machine specifications.
• Achieved higher Area Under the Curve (AUC) values compared to conventional methods on computed radiography and mobile x-ray systems.
• Significantly improved diagnostic performance (p < 0.05) compared to histogram equalization, CLAHE, and unsharp masking.

Conclusions:

• The proposed XM-pipeline significantly enhances AI model diagnostic performance on chest x-rays from diverse machine configurations.
• This pipeline facilitates the widespread clinical application of AI for pulmonary abnormality screening across various x-ray systems.
• The XM-pipeline effectively overcomes the generalization limitations of AI models in medical imaging.