Correction of dose profiles in radiotherapy with the aid of neural networks

Dosimetric measurements in modern radiotherapy with complex therapeutic techniques (e.g., stereotactic radiation therapy) require extensive corrections of the measurement signal, for example, to take disturbing effects due to the presence of a measuring detector into account. Determining these corrections often requires a complex mathematical process. Correcting the so‑called volume averaging effect, for instance, requires the deconvolution of the measured dose profile with regard to the spatial dose response function of the detector.

A newly developed method uses machine learning based on a neural network to deconvolve the dose profile. Determining the detector’s dose response function using a complex procedure is therefore no longer necessary, and the foundations for the routine application of this correction method in hospitals have been laid.

Determining basic data for treatment planning in radiotherapy requires, among other things, the measurement of numerous dose distributions. The accuracy of these measurements is an essential foundation for patient safety and the success of radiotherapy treatment. The characteristics of the detectors used for the measurement must therefore be well known to be able to exclude systematic measurement errors. The increasing application of modern, highly complex therapeutic techniques is thereby continuously raising the requirements made on dose measurements and the dosimetry systems used. Since radiation exposure using these new techniques often drastically deviates from the dosimetric reference conditions, extensive corrections may have to be applied to the measured detector signal to determine the dose.

Very small radiation fields are used especially in the case of stereotactic radiation therapy. These small fields cause a disturbance of the lateral secondary electron equilibrium, which makes accurate measurements of dose profiles very demanding. The presence of the measuring detector causes additional disturbing effects. The signal measured with a dosimeter is, mathematically speaking, a convolution of the undisturbed (“true”) dose profile with the dose response function of the detector used [1]. The dose response function thereby images the detector characteristics as well as their disturbing influence on the measurement signal. The desired “true” dose profile can be determined from the measured signal with the aid of deconvolution methods. For this purpose, established methods such as numerical [2] or analytical deconvolution [3] require an accurate knowledge of the detector’s dose response function which, in turn, is difficult to determine [4].

PTB’s research has led to a new method that uses machine learning [5, 6] based on a neural network (see Figure 1) to deconvolve the dose profile. Determining the detector’s dose response function using a complex procedure is therefore no longer necessary, and the foundations for the routine application of this correction method in hospitals have been laid. In complex photon fields, this new method exhibits a robustness with regard to noise in the measurement data that is superior (see Figure 2) to that of conventional deconvolution methods. Moreover, the method is flexible when faced with the varying spatial resolution of the measurement data and enables quicker profile scanning, which accelerates the workflow in a hospital setting. In addition, this method can be carried out with any probe detector, and it was possible to show that neural networks can also be used with the aid of other radiotherapy and measurement systems. Pre‑trained models of neural networks could thus be provided to hospitals.

To improve the safety and the accuracy of radiation therapy, it is very important to investigate dose profile deconvolution methods that are easy to handle in hospitals in order to guarantee accurate dose measurement even under metrologically challenging conditions. Neural networks have proven to be an efficient and flexible tool to tackle these challenges.

Figure 1:  Architecture of a neural network with a hidden layer for the deconvolution of dose profiles. [7]

Figure 2:  Correction of a dose profile measured with an ionization chamber (crosses). The comparison shows the results of numerical and analytical deconvolution methods as well as the deconvolved profile of the neural network (red curve). The measurement profile of a microDiamond detector (open circles) serves as a reference profile. [8]

References

[1]    Harder D. et al. Convolutions and deconvolutions in radiation dosimetry. In: Comprehensive Biomedical Physics, 2014.

[2]    Looe H.K. et al. Understanding the lateral dose response functions of high‑resolution photon detectors by reverse Monte Carlo and deconvolution analysis. Phys. Med. Biol., 2015.

[3]    Ulrichs A.-B. et al. Entfaltung der durch den Volumeneffekt von Ionisationskammern verzerrten Dosis-Querprofile mit Hilfe der Faltungseigenschaften von Gaußfunktionen. DGMP Jahrestagung (annual meeting), 2016.

[4]    Ketelhut S. and Kapsch R.-P. Measurement of spatial response functions of dosimetric detectors. Phys. Med. Biol. 60 (2015), 6177.

[5]    Liu H. et al. Feasibility of photon beam profile deconvolution using a neural network. Med. Phys. 45 (2018), 5586.

[6]    Schönfeld A.-B. et al. Corrections of photon beam profiles of small fields measured with ionization chambers using a three‑layer neural network. J. Appl. Clin. Med. Phys. 22 (2021), 64.

[7]    Schönfeld A.-B. et al. Applicability of a pre‑trained neural network for the deconvolution of independent ionization chamber‑measured dose profiles in small photon beams. DGMP Jahrestagung (annual meeting), 2021.

[8]    Schönfeld A.-B. Advanced correction methods for dosimetry in complex photon fields. Dissertation, University of Oldenburg, 2021.

Contact

A.-B. Schönfeld, Department 6.2