Physics-informed data augmentation to simulate low dose CT scans: Application to lung nodule detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Convolutional neural networks (CNNs) in medical AI are sensitive to variations from different imaging systems and acquisition parameters.
• These imperceptible image differences can negatively impact AI model performance in clinical applications.

Purpose of the Study:

• To introduce a novel data augmentation technique using imaging physics principles to simulate CT scanner noise characteristics.
• The goal is to improve AI model robustness against variations in radiation dose and acquisition parameters.

Main Methods:

• Physics-Informed Data Augmentation (PIDA) uses CT reconstruction kernel profiles (mAs, Noise Power Spectrum) to simulate dose exposure effects.
• Correlated noise from higher dose scans is inserted into lower dose training data to mimic noise characteristics and enhance variability.
• PIDA was applied to train a lung nodule detection network to mitigate radiation dose-related domain shift.

Main Results:

• PIDA effectively simulated noise characteristics of low-dose CT scans using higher-dose scan data.
• Training a lung nodule detection algorithm with PIDA improved its performance on low-dose CT scans.
• The Competitive Performance Metric (CPM) increased from 0.586 to 0.677 for low-dose scans when PIDA was included in training.

Conclusions:

• PIDA addresses performance degradation in CNNs caused by differences between training and testing datasets.
• The findings demonstrate PIDA's ability to enhance CNN performance for nodule detection on low-dose CT scans despite acquisition variations.