ESTRO 2021 Abstract Book

S1383

ESTRO 2021

Results

The generated pseudoCT images resemble CT and reduce CBCT artefacts as shown in Fig. 1. Quantitative analysis for the 16 test cases, Fig. 2, suggests the AI-based pipeline using the MI algorithm showed a marginal gain in performance in comparison to the direct CBCT-CT pipeline. Better registrations were obtained with longer training. MI DIR showed better robustness to the image quality of the generated images. Conclusion This work reconfirms (p=0.015) that a multimodal DIR algorithm is better suited to CBCT-CT DIR than OF. This advantage was also retained for the pseudoCT-CT DIR pipeline. However, only a marginal improvement was observed when using GANs in H&N CBCT-CT DIR. This should be confirmed on a larger dataset. Refs [1] Isola et al, DOI:10.1109/CVPR.2017.632 [2] Sage et al, Estimating target registration error for automated deformable image registration QA, AAPM 2020 PO-1662 Analysis of different transfer learning approaches when applying AI on small datasets J. Dhont 1 , C. Wolfs 2 , F. Verhaegen 2 1 MAASTRO, GROW – School for Oncology, Maastricht University Medical Centre+, Department of Radiation Oncology, Maastricht, The Netherlands; 2 MAASTRO, GROW – School for Oncology, Maastricht University Medical Centre+, Department of Radiation Oncology , Maastricht, The Netherlands Purpose or Objective Artificial intelligence (AI) has gained widespread application in radiation oncology (RO). However, datasets in RO are typically relatively small, in the order of a few hundreds or occasionally thousands. AI applications in RO are therefore likely to benefit from transfer learning (TL): pre-training the neural network (NN) on a larger source dataset, possibly from another field, and sequentially fine-tuning the NN on the smaller target dataset of the actual task.