Deep Learning-Based Prediction of Daily Delivered Dose for Advanced Lung Cancer Radiotherapy

Abstract

Interfractional anatomical changes reduce the accuracy of the planned radiotherapy treatment. However, developing new plans for every treatment fraction is impractical. A 3D U-Net model was adapted and optimised to predict daily dose distributions in advanced non-small-cell lung cancer (NSCLC), enabling evaluation of dose deviations due to interfractional changes and supporting timely re-planning decisions. The U-Net was trained using planning CT, daily cone-beam CT (CBCT) images, and associated dose distributions, from 24 patients with stage III NSCLC who received 63 Gy in 30 fractions by IMRT (n=13) or VMAT (n=11). Rigid registration was performed to align CBCT images with the corresponding planning CT images, from which dose was re-calculated in the treatment planning system. These CBCT-based dose distributions were assumed to represent the daily delivered (true) dose. Model performance was assessed using a leave-one-out cross-validation (LOOCV) scheme applied separately to three cohorts: IMRT-only (n=13), VMAT-only (n=11), and combined (n=24) patients. Prediction accuracy was quantified using gamma (3%/3mm) analysis, with a 20% maximum dose threshold. The IMRT- and VMAT-trained models achieved mean gamma pass rates of 98.1±1.4% and 90.1±2.6% at 3%/3mm, respectively. The combined-trained models, validated on IMRT and VMAT datasets, attained mean pass rates of 97.3±1.3% and 93.6±2.4% at 3%/3mm, respectively. Gamma failures appeared to correlate with heterogeneous tissue regions, steep dose gradients, and anatomical changes, particularly within the planning target volume (PTV). The higher prediction accuracy of IMRT may be attributable to its fixed beam angles, compared to the continuously rotating gantry in VMAT delivery. The strong agreement of model predictions with true dose distributions demonstrate the potential of U-Net-based dose prediction as a patient-specific, dosimetric verification tool in advanced NSCLC radiotherapy. Further work is needed to test robustness and generalisability of the U-Net model in predicting absolute doses using larger and more diverse datasets.