ESTRO 2023 - Abstract Book

S995

Digital Posters

ESTRO 2023

We retrospectively constructed a dataset of 12,400 verifed endoscopic images from five university hospitals in South Korea, collected between 2010 and 2020 with the participation of otolaryngologists. To calculate the probability of cancer recurrence using various convolutional neural network (CNN) architectures, several deep learning models were developed. Figure 1. Overview of the development and evaluation of the tumor recurrence prediction algorithm.

Results Of the 12,400 total images, 5576 images related to the tongue were extracted. Among these images, analysis was carried out on subjects diagnosed with tongue cancer at Ajou University. We constructed a fusion dataset, which consists of clinical information (gender, age, pathologic stage, surgical margin status) and oral endoscopic images. It showed better results when using fusion data than simple oal endoscopic images (AUROC 0.862 vs 0.959). The results indicate that the best model was VGG 19 (AUROC 0.959 and AUPRC 0.986). Conclusion The deep learning model developed based on the verifed fusion cancer dataset showed acceptable performance in tongue cancer recurrence prediction.

PO-1242 Patient Reported Outcomes Based on EQ-5D-5L Questionnaires in Head and Neck Cancer Patients

T. Sprave 1 , E. Gkika 2 , A. Grosu 2 , R. Stoian 2

1 University Hospital of Freiburg, Radiation Oncology, Freiburg im Breisgau, Germany; 2 University of Freiburg, Department of Radiation Oncology, Freiburg, Germany Purpose or Objective Health economic comparisons of various therapies are often based on health-related quality of life (HRQOL) using EQ-5D questionnaires within the framework of clinical trials. This real-world study evaluates the patient reported outcomes (PROs)-based HRQOL of head-and-neck (H&N) cancer patients undergoing modern radiotherapy (RT) to reflect PRO trajectories. Materials and Methods All H&N cancer patients treated in our clinic between July 2019 and December 2020 who completed the self-reported validated EQ-5D-5L questionnaire (health state index (HI) and Visual Analog Scale (VAS)) at baseline, end of radiotherapy, and at each respective follow up (FU) were included. Descriptive analysis of clinical and sociodemographic data, the frequency and level of each dimension was conducted. To assess the significance of therapy-induced HRQOL changes within and between the group, a distribution-based approach was used. Results Altogether, 366 participants completed a total of 565 questionnaires. For the whole cohort, HI at baseline was 0.804 (±0.208), at RT completion was 0.830 (±0.162), at the first follow-up was 0.812 (±0.205), and at the second follow-up was 0.769 (±0.224). The respective VAS values were 62.06 (±23,94), 66.73 (±82.20), 63.30 (±22.74), and 65.48 (±23.39). Females showed significantly lower HI values compared to males only at baseline (p=0.034). Significantly lower HI values were also seen in patients with definitive RT as compared to adjuvant RT at baseline (p=0.023), the second follow-up (p=0.047), and the third follow-up (p=0.010). As compared to outpatients, inpatients had significantly lower HI values at RT completion (p=0.017), the second follow-up (p=0.007), and the third follow-up (p=0.031). Subgroup analyses by age (<65 vs. ≥ 65) and smoking status (smokers vs. non-smokers) showed no difference at any time points.