Deep learning fluoroscopy denoising: a beam quality insensitive approach

X-ray fluoroscopy is utilized for real-time device guidance during minimally invasive vascular interventions. To minimize radiation dose, fluoroscopy uses low dose images resulting in increased image noise. Deep learning noise reduction techniques offer the potential to enhance device visibility and reduce image noise. However, training a deep learning denoising algorithm typically requires clean training targets which are difficult to acquire from patients. One approach is to add noise to digital angiography images, which are acquired at a higher dose than fluoroscopy, to simulate fluoroscopy images. However, the added noise needs to account for variations in the x-ray beam quality. This work proposes utilizing a noise model that accounts for variable beam quality. The developed noise model does not require prior knowledge of the digital angiography image acquisition parameters. This work compares the performance of neural networks trained with a variable beam quality noise model and a noise model from a single beam quality. The denoising network that was trained with a noise model that included variable beam quality had significantly higher structural similarity index measure (p < 0.001), signal-to-noise ratios (p < 0.001), and significantly less noise (p < 0.001). The area under the curve of the normalized noise power spectrum was 0.034 for the network trained with a variable beam quality noise model and 0.041 for the network trained with a single beam quality. In conclusion, the developed method offers a training scheme for a deep learning-based fluoroscopy denoising algorithm that is insensitive to changes in the beam quality.