Session Item

Artificial neural networks (NN) are used to deconvolve transverse dose profiles of small radiation fields perturbed by the detector volume effect. This study investigates the robustness of a three-layer NN architecture with regards to varying sampling distances and signal-to-noise ratios (SNR) of the input data.

Transverse profiles were acquired using a Semiflex 3D 31021 and, as reference, a microDiamond detector 60019 (both PTW-Freiburg) for dosimetric field sizes ranging from 0.56 x 0.56 cm² to 4.03 x 4.03 cm² and a constant sampling distance of 0.3 mm. The high resolution and high SNR data have been used to train a reference NN model. Firstly, based on the Nyquist theorem, the maximum required sampling distance for each investigated field size has been derived. Secondly, the Semiflex 3D measurements were down-sampled to different sampling rates higher, equal to, or lower than the Nyquist frequency. In combination with two interpolation methods (linear and spline), individual NN models were trained on data with varying resolution to deconvolve and up-sample the measurements matching that of the high resolution reference profiles. Lastly, the robustness of the reference NN model against measurement noise was studied by adding artificial white Gaussian noise to the Semiflex 3D measurements. Subsequently, NN models were trained with the low SNR data. The results obtained using these retrained NN models were compared to the reference NN model.

The maximum required sampling rate according to the Nyquist theorem was found to increase with the field size. For the case where linear interpolation was used, the retrained NN models show poorer performance even though the profile was sampled at a sampling rate higher than the Nyquist frequency. In contrast, NN models retrained for input data sampled at the Nyquist frequency in combination with spline interpolation show comparable results to that obtained with the reference NN model. Deconvolved profiles using the reference NN model show artefacts related to the noise if the input data has lower SNR than the original data used during the training. However, the NN models retrained using data with the same SNR as the input data produced denoised and deconvolved profiles with good agreement to the reference profiles. Further tests revealed that the same agreement can be achieved if the SNR of the training data is lower than that of the input data.

Retraining robust NN models demonstrated the same performance as the reference NN model using high resolution and high SNR training and input data. Even though the sampling rate has been decreased by a factor of 6 and the SNR reduced by a factor of 2, results of similar quality were achieved. The application of these robust NN models allows for faster beam profile scanning and correction for the detector volume effect, while preserving fine spatial resolution and high SNR of the resulting profiles.