ESTRO 2020 Abstract Book

S890 ESTRO 2020

While the volume changes between 3.4% and 59.7%, the maximum electron density changes between 3.8% and 9.2%. In the profiles obtained, there was no difference in the electron density profile between the fractions for ipsilateral lung, whereas there was a difference between the fractions in the profiles obtained for the tumor. Changes in the maximum dose and maximum dose volume in the tumor volume has been demonstrated in the dose distributions obtained on the CBCT images transferred to the planning system. The maximum dose in the tumor volume can increase 2.7%, while the volume of the maximum dose can increase from 3.6% to 6.4%.

against the original labels associated to the range shift maps. Results The analytical method enables the detection of a perturbation of the calibration curve in the soft tissue region, characterized by a positive or negative mean range shift. Setup errors in the AP directions were identified with a gradient in the AP direction, a high standard deviation and a mean range shift around zero (see Figure 1(b)). Calibration curve errors in bone and adipose tissue regions within the selected error range, as well as setup errors in the IS direction are not detectable with this method since all metrics remained around zero in their range shift maps. The CNN method can correctly detect any type of isolated error, as 100% and 99% of the testing images were correctly classified (Table 1, cases I and II). Furthermore, combinations of calibration curve with setup errors can also be identified, since 73% of the maps were correctly classified and 27% were partly classified (Table 1, case III).

Conclusion Detection of inter-dose and dose-volume changes in lung cancer irradiation is important for predicting adaptive treatment planning for the tumor because changes in these parameters affect the dose distribution. PO-1628 Deconvolution of different range error sources using proton radiography and neural networks C. Seller Oria 1 , G. Guterres Marmitt 1 , S. Brandenburg 2 , S. Both 1 , J.A. Langendijk 1 , A. Knopf 1 , A. Meijers 1 1 University Medical Center Groningen- University of Groningen, Radiation Oncology, Groningen, The Netherlands ; 2 KVI – Center for Advanced Radiation Technology- University of Groningen, Medical Physics, Groningen, The Netherlands Purpose or Objective This study assesses the feasibility to detect different sources of error affecting the proton range, individually and in combination, using proton radiography. Material and Methods A square range probing field was simulated in 16 CTs of head and neck cancer patients (see Figure 1(a)). The depth dose profile of each probe in a discretized detector was extracted. Range shift maps were obtained by comparing an unperturbed reference radiogram against a radiogram in which one or more errors affecting range measurements were introduced. Over and under estimations in the conversion from CT numbers to density were simulated in the soft (S+, S-), bone (B+, B-) and adipose (F+, F-) tissue regions. Setup errors in the anterior – posterior (AP+, AP-) and inferior – superior (IS+, IS-) directions were simulated (see Figure 1(a)). Furthermore, the combined impact of calibration curve and setup errors was investigated. First, an analytical method employing mean range shifts, standard deviations and mean gradients in the AP and IS directions as evaluation metrics was used to interpret the range shift maps. Secondly, a convolutional neural network (CNN) was trained to classify the range shift maps into various categories according to each source of error. The CNN was first trained with range shift maps arising from individual sources of error (case I in Table 1), and afterwards including also combinations of errors (cases II and III in Table 1). The performance of the CNN was evaluated for 3 testing sets of maps (I, II and II in Table 1). For each case, the percentage of maps that were correctly, partly and wrongly classified (exact, partial and wrong match) were obtained, comparing the predictions

Conclusion The feasibility of the range probing method within head and neck patients was demonstrated for individual and combined sources of error affecting range accuracy. This outcome provides means of range accuracy quality control in the proton treatment delivery process, with the aim to trigger decisions on plan adaptations.