WindowNet: Learnable Windows for Chest X-ray Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Public chest X-ray (CXR) datasets are often compressed, potentially obscuring diagnostic details.
• Radiologists use windowing on uncompressed images to highlight subtle features, a practice proven beneficial for CT scans but not fully explored for CXRs.

Purpose of the Study:

• To investigate the impact of windowing on the classification performance of machine learning models using chest X-ray (CXR) data.
• To propose and evaluate WindowNet, a novel model designed to learn optimal window settings for CXR analysis.

Main Methods:

• Applied windowing operations to chest X-ray (CXR) images to enhance subtle features.
• Developed WindowNet, a machine learning model capable of learning multiple optimal window settings.
• Compared the performance of WindowNet against a standard architecture without windowing capabilities on the MIMIC dataset.

Main Results:

• Windowing significantly improved the classification performance of machine learning models on chest X-ray (CXR) data.
• WindowNet achieved an average AUC score of 0.812.
• The model without windowing capabilities achieved an average AUC score of 0.759 on the same dataset.

Conclusions:

• Windowing is a crucial operation for enhancing chest X-ray (CXR) classification performance in machine learning.
• WindowNet demonstrates the effectiveness of learning optimal window settings for improved diagnostic accuracy in medical imaging.