Generalizability of deep learning-based dose conversion model in proton beam therapy

In proton beam therapy (PBT), the analytical pencil beam (PB) algorithm involves dose uncertainties in inhomogeneous regions, making accurate Monte Carlo (MC) dose calculation desirable but time-consuming. Deep learning, converting the dose calculated by the PB algorithm into an MC-equivalent dose distribution, can resolve the trade-off between calculation accuracy and speed. Training a DL-based dose conversion model that can be applied to any tumor site would be ideal; however, the appropriate training regions and its generalizability remain unclear.

We developed a DL-based dose conversion model trained on four representative tumor sites (i.e., head and neck, lung, liver, and prostate), and evaluated its generalizability.

Data from 339 patients (a total of 1147 beams) were used. PB doses were obtained from the treatment planning system, and MC doses were calculated using an in-house MC platform. Our developed DL-based dose conversion model was designed to input a treatment planning computed tomography image and PB dose in a single field and output an MC-equivalent dose. The model's generalizability was evaluated on untrained tumor sites, including the esophagus, pancreas, colorectum, brain, breast, cervix, and limb bone and soft tissue. The conversion performance was assessed using 3D γ-analysis and the Dice similarity coefficient (DSC) for isodose volumes.

For most untrained tumor sites, the model achieved average γ-passing rates of ≥90% with a criterion of 3%/2 mm. The esophagus, breast, which are close to the lung, and limb bone and soft tissue showed slightly lower passing rates of 91.3%, 85.9%, and 89.1%, respectively. The average DSC values exceeded 0.8 for most untrained tumor sites.

The proposed DL-based dose conversion model demonstrated high accuracy and generalizability, even for untrained tumor sites. These findings suggest that the model can be adapted to biases in collecting disease data at each PBT center and for rare diseases.