Prompt-gamma imaging for prostate cancer proton therapy: CNN-based detection of anatomical changes

Purpose & Objective
A clinical study (PRIMA) regarding the potential of range verification in proton therapy by prompt-gamma imaging (PGI) is carried out at our institution. As a step towards the automatic evaluation of the measured PGI data, we present an approach to detect anatomical changes in prostate cancer patients from realistically simulated PGI data using convolutional neural networks (CNNs).

Materials & Methods
In-room control CTs (cCTs) were acquired in treatment position before monitoring 146 field deliveries of 10 hypo-fractioned (3Gy/fraction) prostate cancer patients with a PGI slit camera (range: 8-18 fields/patient). After manual CT registration and dose recalculation, spot-wise shifts of integrated depth-dose (IDD) profiles between cCTs and planning CTs were extracted at the 80% distal falloff position and used for ground-truth classification. Treatment fields were considered to be affected by relevant anatomical changes of the patient if >0.1% of all spots (with at least 0.1% of the total monitor units per field) had absolute IDD shifts above 5 mm. These parameters lead to a field-wise IDD ground-truth classification in optimal agreement with a prior manual field-wise classification based on dose difference maps.
Based on the cCTs, we simulated realistic PGI profiles, including Poisson noise and a positioning uncertainty of the PGI slit camera, and extracted spot-wise range shifts by comparison with the expected reference profiles for the planning CT. Spots with reliable PGI information (inside field-of-view and >5E7 protons), were considered with their Bragg peak position for generating two independent 3D spatial maps of 161616 voxels (0.740.740.66 cm3): (1) The PGI-determined range shift in each voxel is the weighted average taking the spot-wise proton number into account. (2) The proton number in each voxel is summed over all respective spots and normalized per field (Fig. 1).
With these maps and the IDD classification, 3D-CNNs (6 convolutional & 2 downsampling layers) were trained using patient-wise 10-fold cross-validation on the binary task to detect anatomical changes.

Results
The CNNs achieved a mean training and validation accuracy of 0.85 (range: 0.77-0.91) and 0.83 (0.70-0.93), respectively (Fig. 2). Based on the validation results, anatomical changes were detected with a sensitivity of 0.88 and a specificity of 0.76.

Conclusion
Our work shows that CNNs can reliably detect anatomical changes in prostate cancer patients from realistically simulated PGI data of clinical irradiations. While a validation on measured PGI data is the next step, this study highlights the potential of an automatic interpretation of PGI data, which is highly desired for routine clinical application and required for the inclusion of PGI in an automated feedback loop for online adaptive proton therapy.