ESTRO 2020 Abstract book

S890 ESTRO 2020

PO-1552 Modelling the increase in local control seen when employing synchronous chemo and IMRT in NPC as BED A. Mel 1 , A. Hartley 2 , J. Assender 3 1 Papua New Guinea National Cancer Centre, Clinical Oncology, Lae, Papua New Guinea ; 2 University Hospitals Birmingham, Oncology, Birmingham, United Kingdom ; 3 University of Birmingham, Institute of Cancer and Genomic Sciences, Birmingham, United Kingdom Purpose or Objective Intensity-modulated radiotherapy (IMRT) is the current standard of treatment for nasopharyngeal carcinoma (NPC), with the addition of synchronous chemotherapy to IMRT being routine practice for the treatment of stage III to IVB NPCs. It was the purpose of this study to assess their relative contribution to the achievement of local control in terms of biologically effective dose (BED) in NPC, utilising similar studies employed in squamous cell carcinoma of the head and neck (SCCHN) as comparisons. Material and Methods Three groups of prospective trials were identified using a Medline literature search. Firstly, those randomising different doses of radiotherapy in the absence of synchronous chemotherapy were identified to calculate the dose-gradient for the BED response to radiotherapy. Secondly, those randomising between radiotherapy alone and synchronous chemotherapy and thirdly, those randomising between IMRT and 2-dimensional conventional radiotherapy (2D-RT) were identified in order to calculate the contributions in BED from these interventions. Trials were included if there was sufficient detail to enable calculation of BED and if 5-year local control was reported. The correlation between the calculated percentage change in tumour BED (tBED) for each method and observed percentage change in local control was also assessed. Results For a value of = 0.3 Gy-1, estimations for the dose- gradient for the BED response to radiotherapy was 0.9, the additional contribution from chemotherapy to local control was equivalent to 2.2 Gy10 and the additional contribution from IMRT to local control was 4.9 Gy10. A strong correlation was observed between the percentage difference in tBED and the percentage difference in local control (P = 0.00001). Conclusion This study has evaluated the role of synchronous chemotherapy and IMRT in NPC in terms of the contribution of each in relation to BED and how these equate to not only local control but also published survival data on the contribution of each intervention to treatment outcomes in NPC. PO-1553 Development of a CT based radiomic model predictive of non-response to chemotherapy in larynx cancer I. Masson 1 , R. Da-ano 1 , F. Lucia 2 , M. Doré 3 , J. Castelli 4 , C. Goislard de Monsabert 4 , J. Ramée 5 , S. Sellami 6 , D. Visvikis 1 , M. Hatt 1 , U. Schick 7 1 LaTIM INSERM UMR 1101 Laboratory of Medical Information Processing, ACTION team Therapeutic ACTion guided by multimodality Imaging in Oncology, Brest, France ; 2 University Hospital of Brest, Radiation Oncology Department, Brest, France ; 3 Institut de Cancérologie de l'Ouest René Gauducheau, Department of Radiation Oncology, Saint Herblain, France ; 4 Institute Eugène Marquis, Radiotherapy Department Cancer, Rennes, France ; 5 Centre Hospitalier de Vendée, Department of Medical Oncology, La Roche sur Yon, France ; 6 Centre Hospitalier de Cornouaille, Radiotherapy Department, Quimper, France ; 7 University Hospital of Brest / LaTIM INSERM UMR 1101 Laboratory of Medical Information Processing, Radiation Oncology Department / ACTION team, Brest, France

patients as below or above the median overall survival (OS)). This second network consists of a 2D CNN exploiting 2 sets (the low dose CT and the PET) of 8×8 tiles (each 64×64 pixels) as 2D inputs representing the entire 3D region containing the tumor in the multimodal PET/CT domain, as training data. The proposed fully automated workflow was first internally evaluated in the first cohort (n=110) and then externally validated in the second cohort (n=37).

Results In the internal validation (n=110), the proposed workflow was able to correctly detect the entire primary tumor in 70% of patients, detected only parts of the tumor in 20%, and failed completely for the remaining 10%. The best accuracy for outcome prediction for patients with tumor correctly detected by the first network was 70%. When applied to the external testing cohort, tumor detection rates were 75%, 5% and 20% for full, partial and missed respectively. Accuracy in predicting outcome was 73% for the patients with whole tumor detection.

Conclusion The availability of large datasets is usually a crucial requirement for efficient training of CNNs. Our work shows that by relying on a two-step approach and transfer learning, these requirements can be in part alleviated. Experiments are ongoing to further improve the accuracy of the prediction, as well the tumor detection.