Range verification in proton therapy: Feasibility of CNN-based detection and classification of treatment deviations from realistic prompt-gamma-imaging data

Introduction
Within a clinical study, we investigate the potential benefit of prompt-gamma-imaging (PGI) based range verification in proton therapy. As the manual interpretation of detected spot-wise range-shift information is time-consuming and complex, we aim to automatically detect and classify treatment deviations from realistic PGI data using convolutional neural networks (CNNs).

Materials & Methods
For 12 head-and-neck cancer patients and an anthropomorphic head phantom, monitoring of single fields from pencil-beam-scanning plans with the IBA slit camera was considered. In total, 386 treatment deviations were simulated on planning and control CTs and manually classified into 7 classes: non-relevant changes (NRC) and relevant changes triggering treatment intervention due to range-prediction errors (±RPE), setup errors in beam direction (±SE), anatomical changes (AC), or a combination of such errors (CE). The spatial maps of filtered PGI-determined range deviations were converted to 16x16x16 voxel grids. Three complexity levels were investigated using 3D-CNNs [training cohort (n=9), test cohort (n=4), Fig.1]: (A) optimal PGI data, (B) realistic PGI data with simulated Poisson noise based on the locally delivered proton number, (C) realistic PGI data with additional positioning uncertainty of the slit camera.

Results
During validation on the independent test data, the 3D-CNNs achieved multi-class accuracies of 81%, 77%, 76% and binary accuracies of 97%, 95%, 93% for the respective complexity levels (A,B,C) (Fig.2). In the most realistic scenario (C), relevant treatment deviations were detected with 97% sensitivity and 82% specificity. Misclassifications of the AC class were caused by similar PGI characteristics of the CE class.

Conclusion
CNNs can reliably detect and classify relevant treatment deviations from realistically simulated PGI data. While validation on measured patient data is needed, our study highlights the potential of automated PGI interpretation, which is desired for broad clinical application and a prerequisite for including PGI in an automated feedback loop for online adaptation.