3247 Introduction: Despite being considered the gold-standard for dosimetry calculations, Monte Carlo simulation (MCS) is not used in clinical practice due to the complexity of its set-up and the long computation time. Therefore deep learning-based simulations were trained to reproduce the MCS, with a substantial reduction in computation time. In particular, Deep-Dose [Lee2019] was proposed to generate 3D dose distributions for internal dosimetry using CT and PET as inputs. However, it uses a sub-optimal geometry description [DeBenetti2022]. We reproduce the concept of Deep-Dose using 3 different Neural Network (NN) architectures introducing the cropping template of De Benetti et al. We evaluate the NN performance in terms of Isodose Consistency [Tan2021] for different numbers of training samples and in terms of generalization.Methods: The Ga-68-PSMA PET/CT of one individual receiving a Lu-177-PSMA for the treatment of prostate cancer were used. PET and CT were resampled to 0.7x0.7x3 mm3. CTs were cropped in patches of size 40x40x16 voxels with 8, 8, and 2 voxels of overlap respectively. The PET patches had size 32x32x14 without overlap, and were zero-padded to match the size of the CT patches. The output of the MCS, performed on the GATE platform [Sarrut2014], was used as ground truth. The MCS physics included beta minus emission, gamma emission, secondary particles tracking, photoelectric effect, different types of scattering, bremsstrahlung and pair production. The CT was converted into materials and densities using the Schneider2000 method [Schneider2000]. 70 million particles were simulated per patch, requiring approximately 75 minutes. In total 221 patches were simulated for this patient. The Lu-177 source was defined with the UserSpectrum [DoseRadar2009].3D versions of U-Net [Çiçek2016], upon which Deep-Dose is built, Tiramisu [Jégou2017] and a custom context-aware Tiramisu (CATNet) were trained to estimate 3D dose distribution as simulated using GATE, receiving the CT and PET patches as a two-channel input. In CATNet, we introduce a context arm, which receives a larger patch of the (unpadded) PET and CT data (80x80x32, downsampled to 40x40x16), that feeds into the bottleneck to provide a broader context of the patch. As training loss, the absolute error voxel error was used. We used the Isodose Consistency [Tan2021] as evaluation metric. The three NN were trained using patches with HU strictly larger than -950, using the lack of Isodose Consistency improvement over 50 epochs as an early stopping criterion. We also performed a 5-fold cross validation. To evaluate generalization power, we used a test set of a different patient (177 patches). The NN were implemented in PyTorch and trained on an NVIDIA TITAN V GPU.Results: All reported values are averages of the validation sets over the 5 folds. In terms of training time, both Tiramisu architectures (standard and context-aware) trained for fewer epochs than the U-Net (204, 174 vs. 285). Also, the Tiramisu architectures resulted in better Isodose Consistency with less training samples (0.653, 0.652 vs. 0.540 for 140 patches). This difference vanishes as the number of training samples increases (0.604, 0.604 vs. 0.583 for 221 patches). In terms of generalization, the CATNet deemed better results both when trained with 140 or 221 samples (0.489 and 0.483, in comparison to 0.472 and 0323 for the standard Tiramisu and 0.384 and 0.378 for the U-Net).Conclusions: For the cropping scheme described by De Benetti et al, the Tiramisu architectures achieve higher Isodose Consistency with less training samples and shorter training periods than U-Net 3D. The CATNet seems to generalize better than the other 2 architectures as the information of the surrounding tissue and uptake may have an impact in the inference of the dose distribution. This effect should be investigated for other isotopes where higher Bremsstrahlung or gamma emissions could have a stronger impact.
«
3247 Introduction: Despite being considered the gold-standard for dosimetry calculations, Monte Carlo simulation (MCS) is not used in clinical practice due to the complexity of its set-up and the long computation time. Therefore deep learning-based simulations were trained to reproduce the MCS, with a substantial reduction in computation time. In particular, Deep-Dose [Lee2019] was proposed to generate 3D dose distributions for internal dosimetry using CT and PET as inputs. However, it uses a su...
»
Login
BibTeX