Physics-informed data augmentation to simulate low dose CT scans: Application to lung nodule detection.
Authors
Affiliations (1)
Affiliations (1)
- Center for Devices and Radiological Health, U.S. Food and Drug Administration, Silver Spring, Marryland, USA.
Abstract
Convolutional neural networks (CNNs) can be sensitive to slight changes in input images, even when the differences are imperceptible to human observers. In artificial intelligence (AI) applications for medical imaging, these variations can result from different imaging systems and acquisition parameters. In this paper, we propose a new approach based on imaging physics principles for simulating the noise characteristics of images from specific CT scanners as part of data augmentation for AI training. Our proposed Physics-Informed Data Augmentation (PIDA) method leverages the mAs and Noise Power Spectrum (NPS) profiles of various CT reconstruction kernels to simulate the effects of various dose exposures. In this approach, the NPS of a higher dose CT scan is used to generate correlated noise, which is then stochastically inserted into the training data. This simulates the noise characteristics of the lower dose exposure and enhances variability within the training set. To demonstrate PIDA's applicability in mitigating radiation dose-related domain shift, we applied PIDA in training a neural network designed to reduce false positives in a lung nodule detection algorithm. We evaluated the impact of the noise insertion training method by assessing lung nodule detection performance on low-dose CT scans. Our experimental results illustrate the effectiveness of our PIDA method in simulating noise characteristics of low dose CT scans from higher dose CT scans. Including PIDA in algorithm training, improved the performance of the algorithm when it was applied to low dose CT scans. The performance in terms of Competitive Performance Metric (CPM) for the low dose scans improved to 0.677 from a CPM = 0.586 when training was performed without PIDA. PIDA is designed to address the performance drop in CNNs due to acquisitional differences between training and testing datasets. Our findings indicate that it enhances the performance of a CNN in detecting nodules on low-dose CT scans when acquisition differences are present.