Training with Real CBCT Data

[SusanneSchaub]

Hello,
Thank you for sharing your code.
I was wondering if this repository has ever been used with real CBCT projection data. In your publication, I read that the model was trained with real data from fluoroscopic images obtained from a GE C-arm Medical System, but was it ever used with real CBCT projections?

I encountered issues when training with real CBCT images using this model. The image quality isn't as good as when training with synthetic projections.

[Ruyi-Zha]

Hi, yes, we trained NAF on real data, and it performed well. However, a slight performance drop is expected due to factors such as noise in real X-ray images, anisotropic effects, and calibration errors. If the performance is significantly bad, I recommend checking for potential issues with the data or scanner parameters and trying tuning hyperparameters. Additionally, you can refer to our latest work, where we also test NAF on the real-world FIPS dataset.

[SusanneSchaub]

Hi, thank you for the quick response. The performance does indeed significantly drop for me when training with real data. I was wondering if the nature of CBCT images could also be an issue. Since CBCT images can have significant variations in gray values compared to CT images, I assume this might make it much harder for the model to learn effectively. You only provided CT images—would it be possible to provide the real CBCT data that you mentioned?

[Ruyi-Zha]

We cannot provide the data used in the NAF paper due to privacy concerns. However, you can use FIPS, which is a public real-world CBCT dataset used in our latest paper, the download link is here. We have converted it to NAF format so you should be able to directly train NAF with the data.

NAF actually works well on FIPS, which shows its effectiveness in handling real-world data. I recommend double-checking the dataset. For instance, verify whether you’ve applied the log operation to convert raw data (white background, black tissue) into integrated data (black background, white tissue). Additionally, confirm that the scanner parameters are accurate.