Session Item

In order to improve the current radiotherapy workflow, MR imaging is proposed as a reference instead of gold standard CT. Indeed, MR allows a better delineation of at-risk-organs thanks to a good soft-tissue contrast. The main drawback of MR is the lack of information about electronic density of tissues which is essential for dose calculation. To face this issue, synthetic CT (sCT) generation from MR is proposed to take advantage of MR accuracy and electronic density information. Moreover, sCT generation is so far dependent on acquisition devices and prevent from its application to all care centers. In this study, we propose a multicentric sCT generation to obtain a generalizable model.

In this study, sixty-nine prostate cancer patients CT and T2-MR were acquired in treatment position. They are from two different centers: 39 patients received a MR imaging with a 3T acquisition device and 30 on a 1.5T acquisition device. All MR were preprocessed to correct their non-uniformity with a N4 bias field correction, a histogram equalization and a filtering by gradient anisotropic diffusion. For the second center, bladders are injected on CTs: a density assignation of 0HU (reference value) was applied to the bladder in CT for each patient.
To generate sCT, 2D cycle-GAN was used, using two ResNet 9blocks as generators and two 70*70 patch-GANs as discriminators. The perceptual loss was computed to compare sCT to CT and the Binary Cross Entropy (BCE) as adversarial loss to classify sCT as real CT or “fake”. Perceptual loss is based on a pre-trained network (VGG16) and 4 layers are used for style and one for content.
Evaluation was performed on a cross validation with 20 patients in the training cohort and 10 patients in the validation cohort. For the multicentric study, a training cohort containing 10 patients from the first center and ten patients from the second center was used.
To evaluate the accuracy of our sCT, mean absolute error (MAE), mean error (ME) and peak signal to noise ratio (PSNR) were computed.

Table 1 shows MAE, ME and PSNR for each training cohort (monocentric and multicentric). No significant differences were found between monocentric trainings and multicentric trainings where some of the test and training data come from the same center.
Figure 1 shows some sCT results from the second center after three different types of training.

To conclude, multicentric context doesn’t penalize trainings when both centers are in the training cohort. The learning cohort should include data from various acquisition devices to generalize our deep learning approach for use in all cancer centers.