Session Item

In this systematic study, we investigate the sensitivity of prompt-gamma imaging (PGI) towards the field-wise detection of inter-fractional anatomical changes in proton therapy (PT) of head and neck cancer (HNC) patients.

Spot-wise range shifts ∆R_PGI were monitored with a PGI-slit-camera during 22 field deliveries of HNC pencil beam scanning (PBS) treatments of 4 patients (field-wise dose per fraction: 0.7-1.0GyE). In-room CTs were acquired for all monitored fractions and range shifts ∆R_IDD at the 80% falloff of spot-wise integrated depth-dose (IDD) profiles served as input for an automatic field-wise ground truth classification (Fig.1). To receive results consistent with an additional manual dose-based classification per field, a PBS spot with relative weight to the field >0.1% was rated as relevant if |∆R_IDD| was ≥5mm for that and at least 1 neighboring spot. Subsequently, a field was classified as relevantly changed if at least 1.5% of all spots were rated relevant.

For the independent PGI evaluation, spots were clustered based on Bragg-peak position and proton number to mitigate statistical measurement uncertainty. Clusters with |∆R_PGI|≥5mm were classified as relevant. For training of the field-wise PGI classification model, the number of relevant clusters, that is necessary to classify the whole field as relevantly changed, was optimized with respect to the IDD ground truth classification using a training set of 11 fields. Finally, the classification model was validated on an independent test set (11 fields).

On the level of PGI spot clusters, there is a significant correlation (r_Pearson=0.3, p<0.01) between IDD and PGI range shifts over all 22 monitored fields (Fig.2B). The only moderate correlation is mainly due to statistical uncertainty of the clusters, represented by the mean absolute error between IDD and PGI range shifts of 3.4mm. The correlation might also be affected by difficult range shift determination due to spots traversing highly heterogeneous tissue. Despite the moderate correlation on cluster level, a field-wise classification, the main endpoint in our study, is possible with high detection sensitivity. Resulting from the training, a field was classified as relevantly changed if >12% of PGI clusters are relevant. The final model achieved a sensitivity of 80% (4/5) and a specificity of 67% (4/6) on the test cohort (Fig.2C).