ESTRO 2024 - Abstract Book

S3912

Physics - Image acquisition and processing

ESTRO 2024

To allow a better convergence of the DL models, a series of preprocessing procedures [1] were carried out to standardise the database. First, a thresholding step was performed on CTs between -1000 and 3000 HU and between 0 and 500 for MRIs. Then, CT and MRIs were rigidly registered using Elastix’s library. Finally, to rectify non-uniformities in the MRI images, the following steps were performed: (1) N4 bias field correction, (2) histogram matching and (3) filtering through gradient anisotropic diffusion. To evaluate the performance of the implemented architectures, a 3 Fold cross-validation method was used with 34 patients for the training, 4 patients for the validation and 13 for the test. For the architecture, conditional GANs were used in different contexts: 2D supervised, 3D supervised and 3D unsupervised [1]. For generators, a ResNet 9 block was used and respectively 70*70 and 70*70*70 patch-GANs for 2D and 3D methods. For the two supervised studies, the generator loss was a perceptual loss computed from a pretrained VGG16 network [3]. For the unsupervised study, a new ConvNext perceptual loss was employed [4]. For the three cGAN-based architectures, the Binary Cross Entropy was used as an adversarial loss to classify sCT images as either genuine CT or "fake" ones. As in the literature [1], to evaluate the accuracy of our image synthesis, the mean absolute error (MAE) and the peak signal-to-noise ratio (PSNR) of the HU in the sCTs compared to the HU of reference CTs were computed in the brain and the extended brain contour. To obtain the extended brain contour, an intersection between the body contour and a five centimetre dilatation of the brain mask was performed.

Results:

Table 1 shows MAE and PSNR results for each DL method (2D supervised, 3D supervised, 3D unsupervised). For the brain, best results were obtained with the 2D method. However, as underlined by the extended brain contour results, the error was higher for bone structures: the MAE on this region becomes smaller for the 3D supervised network than for the 2D one. Then, between the two 3D methods, the supervised one presents smaller MAEs and higher PSNR.