ESTRO 2021 Abstract Book

S514

ESTRO 2021

generated, one for the conventional XT doses and one for the PT doses. Both models received the CT scan, prescriptions of the target volume and binary masks of the organs at risk as input data. The Landelijke Protocol (LIPPv2.2) NTCP model was applied on both the real and predicted dose distributions and for both modalities, to estimate the probability of dysphagia and xerostomia. These effects are estimated by measuring dose on parotids, submandibular glands, oral cavity and pharyngeal constrictor muscles. Patient treatment referral was performed based on six criteria displayed in Figure 1.

Figure 1 : Criteria of the LIPPv2 NTCP model. If one is met, the patient is referred to PT.

Results Out of 47 patients common to both databases, 8 were referred to PT when considering the real data. For our test set, we selected 5 patients referred to PT and 5 to XT to evaluate the performance of the method on both cases. The referral decision based on the predicted dose selected 5 patients for PT for the predicted doses, which corresponded to exactly the same 5 patients selected from the real doses. The same observation was true for patients referred to XT. Moreover, as presented in Table 1a and Table 1b, the criteria leading to proton referral are similar (shown in blue). There are two exceptions: for the xerostomia of grad 2 for patient 1 and criteria 6 of patient 4 but the referral decision did not differ.

Conclusion This combination of DL dose prediction and NTCP models shows promising perspectives to develop an automatic and robust decision support tool for physicians. Observed results suggest that DL will help predict