Session Item

Internal target volume (ITV) definition is commonly carried out based on breathing-correlated 4DCT imaging. However, interfractional breathing variability may lead to sub-optimal ITV dimension over the course of SBRT despite applied motion management strategies. This work aims to optimize the ITV definition using deep learning-based prediction of the patient’s breathing amplitude range after the first dose fraction.

The study includes 259 SBRT sessions of 234 patients with lung and liver lesions. For each session, 10-phase 4DCT data were acquired for ITV definition. Patient breathing was recorded with Varian RPM during 4DCT acquisition and during dose delivery in 5 fractions. The proposed ITV optimization approach consists of two steps. (1) Deep learning modeling: a convolutional neural network was applied using breathing curves of the 4DCT and the first fraction (input data) to predict the amplitude range for the following fractions. Corresponding ground truth is the optimal amplitude range. This range is defined retrospectively as coverage of all breathing amplitudes acquired during dose delivery of fractions 2-5 except of extraordinary irregularities (Fig. 1). Additionally, interfractional amplitude variability was quantified by the Euclidean norm of the 4DCT amplitude range (given by 10-phase reference cycle) and the above optimal amplitude range (Fig. 1). For network training and testing, SBRT sessions were split randomly in train (n=191), validation (n=64) and test set (n=4, pre-selected). Model performance is determined by prediction error (Euclidean norm of predicted and retrospectively optimal amplitude range; Fig. 1). (2) ITV re-definition after the first fraction: the initial ITV is adapted according to the patient’s predicted breathing amplitude range by a 4DCT-based correspondence model that correlates external breathing signal and internal tumor motion.

(1) Model performance was 1.7 +/- 1.1 mm (mean prediction error, validation set). Mean interfractional amplitude variability was 3.4 +/- 2.4 mm. 4-fold cross validation obtained similar mean prediction errors, indicating model reliability. (2) Results of ITV re-definition after the first fraction are summarized in Table 1. The test set prediction error was in the range of 0.3 to 4.2 mm and thus greater than for the validation set. However, the test set interfractional amplitude variability was significantly higher compared to the validation set. Despite reduction of initial ITV size, the resulting adapted ITV would have achieved sufficient tumor motion coverage during dose delivery when implemented in the clinics.

Our results support the assumption that 4DCT-based ITV definition may lead to unfavorable ITVs. However, this mainly occurs in cases of adverse interplay of breathing-sensitive tumor motion and interfractional variability of patient breathing. The proposed approach of ITV re-definition reveals potential of significant ITV volume reduction and normal tissue sparing.