ESTRO 2021 Abstract Book

S649

ESTRO 2021

to determine MOSES threshold that predicted for substantial symptomatic QOL. Categorizing CTCAE max grade as less than or ≥ grade 2 and categorizing MOSES above and below to the determined threshold, CTCAE and MOSES were tested for their accuracy in predicting substantial symptomatic QOL with p<0·05 was considered as statistically significant. Results A total of 201/300 symptomatic patients (either on CTCAE or QOL) were included for analysis. CTCAE maximum grade method demonstrated significant correlation with substantial symptomatic QOL for only diarrhea (p<0·0001) and abdominal pain (p=0·008) whereas MOSES correlated with Diarrhoea (p<0·001), abdominal pain (p<0·0001), urinary incontinence (p<0·01) and fatigue (p<0·001).On testing accuracy, MOSES predicted substantial symptomatic QOL equal or better than CTCAE across all symptoms[Table 1]. Cumulative MOSES including all the above toxicities was also calculated for each patient and was equal or better than CTCAE maximum grade method in predicting substantial symptomatic QOL across all functional scales [Table 2].

Table 1.Summary of correlation of “substantial symptomatic symptom specific QOL” with CTCAE Maximum grade and MOSES for different toxicities

Table 2.Accuracy of CTCAE and MOSES/C-MOSES in predicting substantial symptomatic symptom specific/Functional domains QOL

Conclusion Time imputed toxicity scoring correlates better with QOL symptom items and cumulative toxicity burden using MOSES provides good correlation with functional scales.

Poster discussions: Poster discussion 16: Deep-learning for dose prediction and planning

PD-0818 Dose prediction with deep learning: the effect of data quality and quantity in the model’s accuracy A. Barragan 1 , M. Thomas 2,3 , G. Defraene 4 , S. Michiels 5 , K. Haustermans 6,7 , J. Lee 8 , E. Sterpin 8,2 1 UCLouvain, Molecular Imaging Radiation Oncology (MIRO), Brussels, Belgium; 2 KULeuven, Department of Oncology, Leuven, Belgium; 3 University Hospital Leuven, Department of Radiation Oncology, Leuven, Belgium; 4 KULeuven, Department of Oncology, Brussels, Belgium; 5 UCLouvain, Molecular Imaging Radiotherapy and Oncology (MIRO), Brussels, Belgium; 6 KULeuven, Department of Oncology, Leuven, Belgium; 7 University Hospitals Leuven, Department of Radiation Oncology, Leuven, Belgium; 8 UCLouvain, Molecular Imaging Radiotherapy and Oncology (MIRO), Brussels, Belgium Purpose or Objective Radiotherapy dose prediction with deep learning (DL) has recently become a popular topic. Most research has focused on improving the accuracy of these models by modifying their architecture. However, little attention has been paid to the influence of the training database on the model’s performance. The present study aims to switch the attention from architecture to data, by analysing the effect of data quantity and quality on the performance of DL models for dose prediction of intensity-modulated radiotherapy (IMRT) of esophageal cancer. Materials and Methods Two databases were used: a variable database (VarDB) with 56 clinical cases extracted retrospectively,