Adaptive Normalization Enhances the Generalization of Deep Learning Model in Chest X-Ray Classification

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Chest X-ray (CXR) analysis relies heavily on image preprocessing for accurate classification.
• Standard normalization techniques may not optimally handle variations in CXR contrast and acquisition.
• Developing robust preprocessing methods is crucial for reliable AI in medical diagnostics.

Purpose of the Study:

• To benchmark adaptive preprocessing against standard methods for CXR classification.
• To evaluate the impact of histogram standardization and ROI cropping on model performance.
• To assess the computational efficiency and statistical significance of the adaptive pipeline.

Main Methods:

• Controlled benchmarking of min-max scaling, Z-score normalization, and an adaptive pipeline (ROI cropping + histogram standardization).
• Evaluation across four public CXR datasets and three convolutional neural network architectures.
• Statistical analysis using Friedman-Nemenyi and Wilcoxon signed-rank tests.

Main Results:

• The adaptive pipeline generally enhanced accuracy and F1-score, especially on datasets with stable contrast.
• Histogram standardization was the main driver of performance improvement, with ROI cropping providing complementary benefits.
• Minimal computational overhead was observed (+6.3% training time; 5.2 ms/batch).

Conclusions:

• Adaptive normalization, particularly histogram standardization, is a practical preprocessing choice for CXR classification.
• The proposed pipeline improves cross-dataset robustness and reliability in AI-driven CXR analysis.
• The method offers statistically significant performance gains with negligible computational cost.