AI facilitates radiation therapy planning

The more accurate the radiation planning, the easier it is to determine the exact dose that will most effectively fight the tumor while being as gentle as possible to the patient. Various measurements are incorporated into the planning. The demands on these measurements and on the dosimetry systems are escalating as ever more advanced and highly complex treatment techniques are introduced. In many cases, the measurements for determining basic data using different detectors in water phantoms deviate significantly from the dosimetric reference conditions, and extensive corrections of the measured detector signal are sometimes required in order to determine the dose.

Stereotactic irradiation is a highly precise radiotherapeutic method used to treat small tumors or metastases very effectively. It allows the precise application of high individual doses whilst ensuring maximum protection of sensitive organs. However, measuring the dose profiles accurately is very difficult due to the extremely small fields used. Furthermore, the dosimeter probe itself causes interferences. In mathematical terms, the signal measured with the dosimeter is a convolution of the undisturbed (“true”) dose profile with the dose response function of the applied detector. Such established methods as numerical or analytical deconvolution require an exact knowledge of the dose response function of the detector, which is difficult to determine.

A new method applies machine learning: It uses a neural network for the deconvolution of the dose profile. This means that it is no longer necessary to laboriously determine the dose response function of the detector, paving the way to routine clinical use. Compared to classical/conventional deconvolution methods, the new method shows superior robustness in complex photon fields as regards noise in the measurement data. Moreover, the method is flexible in the presence of varying spatial resolution of the measurement data and allows faster profile sampling, which speeds up the clinical workflow. Beyond that, the method may be used with any type of dosimeter probe. It also has been shown that neural networks can be used for other irradiation and measurement systems. Clinics could thus be provided with pre-trained models of neural networks.