# Deep learning‐based upsampling of 2D detector array measurements for patient plan verification in radiotherapy

## TL;DR

This paper introduces a deep learning method to improve the resolution of radiation therapy dose measurements, leading to more accurate patient plan verification.

## Contribution

A novel deep learning approach is proposed to enhance detector array resolution for accurate dose verification in radiotherapy.

## Key findings

- Neural network upsampling increased gamma index passing rates by up to 20% compared to bilinear interpolation.
- VMAT plan passing rates improved by 7-8% using neural network upsampling with TPS dose as reference.
- The method achieved a threefold resolution increase from 5 mm to 1.7 mm for the OCTAVIUS Detector.

## Abstract

Detector arrays are commonly used for treatment plan verifications in intensity modulated radiation therapy. However, the intrinsic resolution of detector arrays is limited by the physical dimensions of each single detector and the detector‐to‐detector distance. This may lead to inaccurate representations of steep gradients and narrow dose peaks.

This work presents a deep learning approach for increasing the effective spatial resolution of detector arrays used for patient plan verification. The presented approach aims to augment missing values in the insensitive areas of the detector matrix and to increase the sampling frequency of the measured 2D dose profile. Furthermore, perturbations caused by finite detector's dimensions via the volume‐averaging effect are corrected during the upsampling process.

In this work, Monte Carlo simulation methods were employed to synthetically generate training data, enabling a wide coverage of different linear accelerator setups and field shapes. The approach was implemented for the OCTAVIUS Detector 1500 (PTW Freiburg, Germany), which consists of 1405 air‐filled ionization chambers arranged in a checkerboard pattern. This arrangement enables a threefold increase in resolution from 5 mm, achieved with the standard bilinear interpolation, to 1.7 mm using neural networks. The implemented neural networks are based on a deep convolutional architecture and were trained using PyTorch. Initially, the models were tested by comparing the upsampled measurements of individual step‐and‐shoot IMRT segments with measurements obtained using a high‐resolution OCTAVIUS Detector 1600 SRS liquid‐filled ionization chamber array. In addition, radiochromic film measurements of fields with leaf gaps of 1 and 2 cm were used to demonstrate the differences between measurements and interpolation results in the presence of steep gradients and narrow dose peaks. Finally, reconstructed 3D dose distributions of VMAT plans, using both the original and upsampled measurements, were compared to the treatment planning system calculations.

The comparison of the individual IMRT segments with a standard bilinear interpolation showed an average increase in the gamma index passing rate of up to 20%. In the case of 3D dose reconstructions from the field‐by‐field IMRT measurements in the OCTAVIUS 4D phantom, the neural network upsampling yielded an average increase in passing rate of 22% as compared to bilinear interpolation, when using the OD 1600SRS array measurements as reference and 19% with the TPS calculated dose distribution as reference. Similarly, for VMAT plans, the passing rate showed an average increase of 8% for the measurement at an Elekta accelerator and 7% at a Varian Ethos accelerator, both using the TPS dose distribution as reference.

It has been shown that a neural network can be applied to upsample the detector array resolution. This results in a better interpolation of measurement points, especially in regions of steep gradients, than compared to a standard bilinear interpolation. The passing rates of all investigated VMAT plans are increased by applying the proposed neural network upsampling approach.