ESTRO 2022 - Abstract Book

S1398

Abstract book

ESTRO 2022

Materials and Methods The goal was to obtain improved resolution and contrast in the same time with equivalent Signal-to-Noise Ratio (SNR). All acquisitions were performed with the patient in treatment position using radiotherapy immobilization devices on a 1.5T Sola MRI (Siemens). T1 and T2 FLAIR acquisitions were investigated for brain lesions and T2 acquisitions for pelvic purposes. Sequences were first optimized on phantom using NEMA subtraction method in order to obtain equivalent SNR, as there were no real guidelines yet for compressed sensing factors and denoising factors. After, patient acquisitions were evaluated quantitatively (post-contrast T1 and T2 FLAIR) by a student t test and qualitatively through a randomized alternative choice test (1=CS preferred, 0 = equal, -1 = CAIPIRINHA) by three expert radiologists and two radiation oncologists through a Wilcoxon signed rank test. Rater agreement was evaluated by a pairwise Kappa Cohen test. Results CS 3D T1 brain acquisitions were evaluated as superior both quantitatively and qualitatively (rater value = 0.6, p < 0.05) and benefited from improved lesion contrast of 7% (p=0.017, 17 lesions) due to reduced repetition time which would otherwise lead to insufficient SNR for CAIPIRINHA acceleration. T2 FLAIR acquisitions were quantitatively superior (resolution), but qualitatively evaluated as equivalent with rater value 0.3 (p > 0.05). 3D T2 pelvic acquisitions were evaluated as inferior with CS acceleration: no significant gain in resolution/SNR was obtained and mean rater value was - 0.5 (p<0.05). The results of the raters suffered however from large interrater variability with pairwise Cohen’s Kappa < 0.33. Conclusion Compressed sensing is an useful technique to overcome the flexible coil SNR issues for radiotherapy preparation for intracranial T1 SPACE and T2 FLAIR. However for pelvic 3D T2 images, k-space based acceleration techniques should be used. J. Nunes 1,1 , S. Fettem 1 , S. Tahri 1 , L. Macke 1 , H. Chourak 2 , A. Barateau 1 , C. Lafond 1 , R. de Crevoisier 1 , I. Bessieres 3 , L. Marage 3 , O. Acosta 1 1 Université de Rennes 1, LTSI (Laboratoire du Traitement du Signal et de l'Image), INSERM UMR 1099, CLCC Eugène Marquis , Rennes, France; 2 Université de Rennes 1, LTSI (Laboratoire du Traitement du Signal et de l'Image), INSERM UMR 1099, CLCC Eugène Marquis, Rennes, France; 3 Centre Georges-François Leclerc (CGFL), Departement of Medical Physics, Dijon, France Purpose or Objective In the context of MR-only radiotherapy workflow, several deep learning methods (DLMs) have been developed for synthetic- CT (sCT) generation from MR images. The Pix2Pix DLM (a conditional generative adversarial network [cGAN]) can be applied on the 3 MRI views (transverse, sagittal and coronal) and not only on the axial view. The aim of this study was to compare the sCTs resulting from the 2D+ Pix2Pix model (in the 3 views) and the 2D Pix2Pix model (axial view) for prostate MRI-only radiotherapy. Materials and Methods Prostate CT and MR images were acquired in treatment position for 39 patients. MR acquisitions, using T2/T1-weighted TrueFISP sequences, were performed with an MRI-linac device (MRIdian, Viewray, 0.35T). 2D+ method consists of generating 3 sCTs (according to each view) per patient, and combined in one sCT by using the median voxel value. sCTs generated by the 2D Pix2Pix model (axial view) were compared to sCTs generated with the 2D+ Pix2Pix model. For both of these methods, the perceptual loss function, a ResNet 9 blocks generator, a PatchGAN discriminator, and Adam optimizer were used. The evaluation was performed on a 5-fold cross validation using 30 patient images for training and 9. Finally, both sCT were compared to the original CT from a voxel-wise comparison with the mean absolute error (MAE) in Hounsfield units (HU), mean absolute percentage error (MAPE) in % , and peak signal to noise ratio (PSNR) in dB. The Wilcoxon test was used to compare the results obtained with the 2D+ model to those obtained with the 2D model. Significant differences were considered for p-value<0.05. Results Table 1 presents the results of MAE, MAPE and PSNR for the two methods. For the body and the bones, significantly lower MAE and MAPE results were found with the 2D+ Pix2Pix model, compared to the 2D Pix2Pix model. For the body and the bones, significantly higher PSNR results were found with the 2D+ Pix2Pix model, compared to the 2D Pix2Pix model. sCTs generated from 2D+ Pix2Pix model were less impacted by inter-slices artefacts (Figure 1) than sCTs generated by 2D Pix2Pix model. PO-1611 Evaluation of synthetic-CT generated from prostate MRI (0.35T) with a 2D+ Pix2Pix method

MAE (HU)

MAPE (%)

PSNR (dB)

Body

Bones

Body

Bones

Body

Bones

2D Pix2Pix

34.6 ± 7.1

136.8 ± 20.7 1.2 ± 0.3

0.5 ± 0.1 29.8 ± 1.6 18.7 ± 1.2

29.2 ± 5.0 * 121.0

±

31.0 1.6 *

±

2D+ Pix2Pix 19.1 ± 1.4 * Table 1: MAE, ME and PSNR (mean ± standard deviation) results in the body and the bones for 39 patients *Significant differences were considered at p-value<0.05. 20.4 * 1.1 ± 0.2 * 0.4 ± 0.1 *