Effect of arc length on the deep learning prediction of monitor units in lung stereotactic ablative radiation therapy treatment

Abstract

Introduction

The dose magnitude required to fine-tune radiation in multi-lesion stereotactic ablative radiation therapy (SABR) treatment to the lung is driven by the monitor units (MU) per control point (CP). We investigate the arc length effect on the deep learning (DL) prediction of the MU per CP for automated lung lesions treatment planning.

Methods

Consecutive lung cancer patients treated at our institution between 01/2019 and 11/2024 were considered. Two models were trained, one on a homogeneous (same-nCP) and the other on a heterogeneous (diff-nCP) set of arc lengths with an equivalent number of samples. A third model was trained with an increased sample size of heterogeneous arc lengths (all-nCP). The predicted MU per CP were converted to meterset weights and MU per beam. The dosimetry achieved with predicted MU per CP was compared with the clinical dosimetry using gamma passing rates (γPR) and achieved clinical goals.

Results

In total, 60,720 samples from 295 treatments of 257 patients were included. The mean absolute percentage error between predicted and clinical meterset weights/MU per beam was less than 5.5 %/5.3 % with the all-nCP model and less than 8.3 %/7.1 % with the same-nCP and diff-nCP model. The median γPR(3 %, 2 mm) was 100 % with the all-nCP model and greater than 99.4 % with the same-nCP and diff-nCP models. All models provided the same or greater number of achieved clinical goals.

Conclusions

DL model trained with variable arc lengths allowed increased sample size and provided equivalent dosimetry in multi-lesion SABR treatment to the lung.