ESTRO 2024 - Abstract Book

S3907

Physics - Image acquisition and processing

ESTRO 2024

We have demonstrated the effectiveness of our method through sparse-view CT reconstruction. We employed a dataset of high-quality images of planning fan-beam CT in radiation therapy, consisting of 3775 images from 20 patients. These images served as the training dataset for the Denoising Diffusion Probabilistic Model (DDPM). Subsequently, using projection data from a patient not included in the training set, we reconstructed both full scan CT and sparse-view CT using the following methods:

1) Iterative reconstruction (IR)

2) IR with total variation regularization

3) Only the diffusion model

4) IR combined with the diffusion model (the proposed method)

To quantitatively evaluate image quality, we employed SSIM and PSNR metrics, with the full scan CT image serving as the ground truth. The results of the reconstructed images, along with the corresponding metric values, are presented in Figure 2. While it may appear that method 3) (only the diffusion model) enhances image quality, the metric values are not particularly high. This is primarily due to the fact that method 3) alters the patient's anatomical structures within the image. In contrast, our proposed method successfully improves image quality while preserving these structures and achieves the best results among methods 1) to 4), based on quantitative metrics.

Conclusion:

We combine the diffusion model with iterative CT reconstruction to enhance the image quality of sparse-view CT while preserving the patient's anatomical structures. Our method's performance surpasses that of the traditional CT reconstruction method, specifically iterative reconstruction with total variation regularization. Moreover, because we exclusively employ high-quality CT images during training in an unpaired manner, we can apply this innovative reconstruction method not only to sparse view CT but also to other low-quality CT scans, such as low-dose CT, CBCT, and MVCT, using a single trained model.