ESTRO 2023 - Abstract Book

S1372

Digital Posters

ESTRO 2023

Conclusion This study presents an extensive validation of a pre-market release of the novel CBCT imager. The variety in size, materials and other phantom features allowed qualitative and quantitative evaluations of multiple imaging features. The novel CBCT panel has a large sFoV and additional eFOV capability. Although more pronounced variations were observed for the HU values of tissue-mimicking inserts in the eFoV, geometrical accuracy is comparable to standard planning CT eFoV reconstruction.

PO-1669 Optimizing MRI-based temperature-controlled hyperthermia by reconstructing undersampled acquisitions

R. Khatun 1

1 Universitätsklinikum Erlangen, Department of Radiation Oncology, Erlangen, Germany

Purpose or Objective Hyperthermia in conjunction with radio-and/or chemotherapy has become a recognised cancer treatment for selected tumor entities. For 60 min, the tumour tissue is exogenously heated to temperatures of 40 to 41°C. Using dynamic Magnetic Resonance Imaging (MRI), temperature may be monitored non-invasively. However, a major challenge is that the MRI is an inherently slow process - scan time for high-resolution imaging compromises with the temporal resolution. Undersampling reduces the speed of acquisition by ignoring parts of the data, but results in a loss of resolution. Several deep learning-based methods have already been proposed, but they mainly focus on the magnitude images, while the phase images are ignored which are fundamental requirements for MR thermometry. This work reconstructs magnitude as well as phase images and aims to reconstruct complex images. Materials and Methods This research employs the Fourier-PDUNet and Fourier-PDNet models in the NCC1701 pipeline to reconstruct highly undersampled MR thermometry using complex-valued convolutions – while directly working with the complex images. MRIs from 44 individual patients with various forms of sarcoma cancer, in most cases originating in the patients' legs, were utilised to evaluate this method. The dataset of 48 cancer patients was randomly divided into three sets: training, validation, and testing with 26, 7 and 11 subjects, respectively. The original images were undersampled with an acceleration factor of 4. Results The Structural Similarity Index or SSIM is used as a metric to measure the similarity between two given images. The undersampled images resulted in average SSIM values of 64% and 31% for the magnitude and the phase images, respectively. The Fourier-PDUNet and Fourier-PDNet models managed to reconstruct those data with average SSIM values of 90% and 91% for the magnitude images while achieving 43% and 44% for the phase images, respectively. The root mean square error (RMSE) of the reconstructed temperature map was determined. Here, the temperature difference between the ground truth and the undersampled images was 1.5080 +/- 059, which is 1.5 °C higher than the ground truth. However, the mean RMSE of Fourier-PDUNet is 1.0790 +/-039, and Fourier-PDNet is 1.0790 +/- 036, with a 1°C discrepancy with the ground truth. As a result, the models provide 40% greater accuracy when reconstructing the temperature maps.