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

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Proton beam therapy (PBT) faces dose uncertainties in inhomogeneous regions with analytical pencil beam (PB) algorithms.
• Accurate Monte Carlo (MC) dose calculations are time-consuming, creating a trade-off between speed and precision.
• Deep learning (DL) offers a solution by converting PB doses to MC-equivalent distributions, but generalizability across tumor sites requires investigation.

Purpose of the Study:

• To develop and evaluate a DL-based dose conversion model for PBT.
• To assess the model's generalizability across diverse and previously untrained tumor sites.

Main Methods:

• A DL model was trained on PBT data from 339 patients across four tumor sites (head and neck, lung, liver, prostate).
• The model inputs CT images and PB doses to output MC-equivalent doses.
• Generalizability was tested on seven untrained tumor sites, with performance assessed using 3D γ-analysis and Dice Similarity Coefficient (DSC).

Main Results:

• The DL model achieved high accuracy (≥90% γ-passing rate with 3%/2mm criteria) for most untrained tumor sites.
• Slightly lower passing rates were observed for esophagus (91.3%), breast (85.9%), and limb bone/soft tissue (89.1%).
• Average DSC values exceeded 0.8 for most untrained sites, indicating good performance.

Conclusions:

• The developed DL model demonstrates significant accuracy and generalizability for MC-equivalent dose conversion in PBT.
• The model's adaptability extends to various tumor sites, including those not included in the initial training.
• This approach can aid PBT centers in handling diverse patient data and rare diseases.