Range verification in proton therapy: Can prompt-gamma imaging identify the source of deviation?

Purpose/ Objective
In-vivo prompt-gamma imaging (PGI) is a promising method for directly assessing deviations in the proton range during proton therapy. However, several effects that can cause range shifts in patients need to be distinguished, e.g. global errors in CT conversion to stopping power ratio (SPR), variations in patient setup, and changes in the patient anatomy. Here, we evaluate if the source of range deviation in proton pencil-beam scanning (PBS) can be distinguished based on PGI information using a slit camera [1].
Material and Methods
For a virtual head-and-neck tumor in an anthropomorphic head phantom, a PBS treatment plan with simultaneous integrated boost (3 beams, 70Gy and 57Gy in 33 fractions) was generated. For all PBS spots in the investigated beam, PGI profiles were simulated using a verified analytical model of the slit camera [2, 3] for the reference scenario as well as for different error scenarios: SPR change of ±1.0, ±2.0 and ±3.5%, setup error in beam direction of ±1mm and ±3mm, and 10 scenarios of realistic anatomical changes (Fig. 1). A decision-tree approach was proposed to classify different groups of error sources. This included preceding filtering of PBS spots containing reliable PGI information for range verification. For simplification and better hypothesis generation, the head phantom was first overridden with water density. Afterwards, the real phantom anatomy including all heterogeneities was analyzed. It was evaluated whether the different error scenarios could be classified correctly.

Results
An automated filter to identify reliable PBS spots was developed, e.g. assuring that the spot position is within the effective field of view (FOV) of the camera and that the fall-off of the PGI profile is completely included in the FOV – even in case of range shifts. For subsequent decision-tree-based error source classification (Fig. 2), the following parameters were selected: The coefficient of determination (R2), the slope and intercept of the linear regression between range shift and penetration depth as well as the 2D range shift map. With this approach, 27 of 30 error scenarios could be identified correctly. However, the three error scenarios with anatomical changes in the nasal cavity could not be identified because the automated filtering approach had removed most relevant spots in this region.
Conclusion
An automated classification approach was introduced to identify the source for range deviation solely from prompt-gamma information. Based on phantom data, including simulation of realistic anatomical variation, the results are promising. Further refinement of this initial approach might be beneficial. An extension of the validation with patient CT data is in preparation. In the future, an application of the approach on clinically measured PGI data is planned. Also other classification methods could be evaluated.