# Prescription‑dose stratification improves deep learning‑based VMAT dose prediction in locally advanced NSCLC

## TL;DR

Stratifying prescription doses improves deep learning predictions of radiation therapy doses for lung cancer, enhancing accuracy in target coverage and organ protection.

## Contribution

This study demonstrates that prescription-dose stratification enhances deep learning-based VMAT dose prediction accuracy in locally advanced NSCLC.

## Key findings

- Single-prescription models achieved MAEs < 4 Gy for PTV coverage and < 1 Gy for hot-spot dose metrics.
- Mixed-prescription models showed higher errors, with PTV hot-spot MAE reaching 11.3 Gy and spinal cord dose errors up to 5–6 Gy.
- Voxel-wise differences revealed local deviations in low-dose lung regions and near dose gradients.

## Abstract

Volumetric-modulated arc therapy (VMAT) planning for locally advanced non-small cell lung cancer (NSCLC) is an iterative and planner-dependent process that often requires multiple optimization cycles to balance target coverage and organ‑at‑risk (OAR) sparing. Deep‑learning dose prediction can accelerate planning by providing patient‑specific reference dose distributions, but the impact of prescription‑dose mixing during model training remains unclear. This study evaluated whether prescription‑stratified models improve VMAT dose prediction performance. Seventy-two NSCLC VMAT cases were recalculated to 50, 54, and 60 Gy and split into training, validation, and test sets (42/10/20 cases). Four models with identical 3D U-Net architecture were developed: three single-prescription models (50/54/60 Gy) and one mixed-prescription model (50 + 60 Gy). Performance was assessed using mean absolute error (MAE) for planning target volume (PTV) and OAR dose metrics. Single-prescription models reproduced PTV coverage (D95% and D99%) with MAEs < 4 Gy and hot-spot (D2cc and D5cc) errors < 1 Gy, while mean dose errors for lungs and heart were ≤ 2.3 Gy. The mixed-prescription model showed larger errors: PTV hot-spot MAE rose to 11.3 Gy, and spinal cord maximum-dose errors reached 5–6 Gy, although most other OAR metrics changed modestly. Voxel‑wise difference maps revealed local deviations of a few Gy in low-dose lung regions and near steep gradients. These findings indicate that prescription‑dose stratification improves clinically relevant prediction metrics and support deep‑learning dose prediction as a planning decision‑support and optimization‑guidance tool.

The online version contains supplementary material available at 10.1038/s41598-026-43192-6.