ESTRO 2024 - Abstract Book

S3761

Physics - Image acquisition and processing

ESTRO 2024

As shown in Fig. 1, we established a series of steps to generate high-quality enhanced 4D-MRI. The main procedure entailed the following two steps: 1) setting high-speed multiphase LAVA sequences to acquire low-quality enhanced 4D-MR images while the patient breathed freely. The paired low/high-quality breath-holding (BH) LAVA images were also scanned. 2) To improve the quality of the scanned 4D-MR images, we developed a personalized high-quality enhanced 4D-MRI generation model based on deep learning, which is guided by paired low/high-quality MRI images and trained for each patient to deal with the diversity that exists among patients with liver cancer. As a result, the low-quality 4D-MR images were input into the well-trained personalized model to generate high-quality ones. Fifty eight patients who underwent radiotherapy for liver tumors were included. The patients were randomly assigned to a training dataset (45 patients) and a test dataset (13 patients).

Results:

1. Image quality evaluation

Image quality was compared among the scanning 4D-MRI (low-quality, LQ), the generated 4D-MRI using the general model (GM), and a personalized model (PM). The results of Statistical and intuitionistic comparisons are shown in Table 1. Compared with low-quality 4D-MRI, the overall SSIM values of personalized generated images were improved from 80.77 ± 3.86 to 83.94 ± 3.15 for the BH dataset and from 79.07±4.04 to 81.60±3.84 for the FB dataset. Furthermore, the results of the SSIM, PSNR, and MSE also demonstrate the substantial image quality improvement that came with the use of a personalized strategy compared to the use of the general DL model. The p-value for the proposed method is less than 0.001, indicating highly significant improvements. Our method improves image quality while preserving anatomical structures.