ESTRO 2023 - Abstract Book

S1911

Digital Posters

ESTRO 2023

Conclusion Incorporating FLL into severe lymphopenia probability models improved model quality with modest changes in discrimination capability. A larger cohort is needed to conclude these findings.

PO-2123 Deep Learning Prediction of 2-year Survival in Lung Cancer Patients Undergoing Radical Radiotherapy

Y. ZHANG 1 , M. Simard 2 , O. Chohan 3,4 , Z. Shen 5 , P. Zaffino 6 , G. Royle 3 , D. Brand 3,7 , M. Hawkins 3,7 , C.A. Collins Fekete 3

1 University College London, Department of Medical Physics and Biomedical Engineering, London, United Kingdom; 2 University College London , Department of Medical Physics and Biomedical Engineering, London, United Kingdom; 3 University College London, Department of Medical Physics and Biomedical Engineering, LONDON, United Kingdom; 4 University College London Hospitals NHS Foundation Trust, Department of Radiotherapy Physics, LONDON, United Kingdom; 5 University College London, Medical Physics and Biomedical Engineering, LONDON, United Kingdom; 6 Magna Graecia University of Catanzaro, Department of Experimental and Clinical Medicine , Catanzaro, Italy; 7 University College London Hospitals NHS Foundation Trust, Department of Radiotherapy, LONDON, United Kingdom Purpose or Objective Non-small cell lung cancer (NSCLC) is a cancer of unmet need with poor prognosis, as illustrated by the low 5-year overall survival (OS) of 16.6% in the UK. Dose escalation trials (e.g. RTOG 0617) have failed to improve trial-level OS. However, some individual patients might derive benefits from dose escalation. We aimed to train a model to predict 2-year OS, including standard and dose-escalated patients. Materials and Methods An AI classification model was trained on the RTOG-0617 dataset (n=412 NSCLC patients) containing: clinical data, planning CT, planning contours and radiotherapy dose distribution. Patient survival was 0.5 to 61.5 months. The dataset is balanced for 2-year OS: 182 patients (44%) survived for 2 or more years, and 230 (56%) survived less. A multi-channel 3D DenseNet (Figure 1) was used to extract relevant features from CT images, PTV contour and dose distributions. The extracted image features were appended to 26 clinical features plus dose metrics from organs at risk (OARs) and passed to a feed-forward neural network for classification. The proportion of the training, validation and holdout dataset is 70%, 15%, and 15%, respectively. Bootstrapping was performed (30 epochs per training, 15 bootstrap samples) to obtain prediction statistics. We applied ensemble voting to the top five bootstrap models (based on AUC).