Differentiable Autoplanning Algorithm for Proton Therapy Based on Deep Learning: A Proof-of-Concept

Proton therapy provides improved preservation of healthy tissues and dose conformity compared to conventional radiotherapy; however, current planning practices remain limited by manual, iterative optimization procedures. This study introduces a novel proof-of-concept approach leveraging a deep learning (DL)-based differentiable dose engine for comprehensive and automated proton therapy planning. Utilizing a Graph Neural Network (GNN) architecture optimized for cylindrical geometry, the developed model achieves rapid and accurate dose computations, enabling gradient-based optimization of all relevant beam parameters, including fluence, energy layers, spacing, and field orientations. Validation performed on simplified homogeneous scenarios demonstrates the feasibility of achieving uniform and precise dose distributions, efficiently converging to clinically relevant solutions within milliseconds. These promising results highlight the potential of integrating DL-driven differentiable dose engines into clinical workflows, paving the way for realtime optimization, robust planning methodologies, and advanced proton delivery techniques.