A Clinically Scalable AI-Based Deformable Image Registration Framework for Planning-to-Repeat CT Alignment In Proton Therapy
Abstract
Purpose
In proton therapy, misalignment between planning and repeat CT arises from variations in patient positioning, physiological motion, and treatment-related anatomical changes, compromising contour propagation and treatment plan adaptation. Conventional deformable image registration (DIR) methods are often slow and less robust under large anatomical variations. Although AI-based DIR methods have shown promise, most rely on idealized assumptions and limited training data. This work aims to develop a robust, clinically scalable DIR framework for proton therapy using large-scale, heterogeneous clinical data.
Methods
We developed an AI-based foundation deformable registration model for clinical CT alignment in proton therapy using a transformer-based progressive coarse-to-fine strategy to handle large anatomical variations. Clinically available planning information, including organ contours, dose maps, and treatment plans, is incorporated through innovative strategies such as foreground-aware optimization, anatomy- and risk-guided attention, and text-conditioned clinical priors. The model was trained on a large-scale proton therapy dataset collected at our institution since 2016, covering diverse anatomical sites, disease types, patient populations, and imaging standards.
Results
The proposed method was evaluated on 1,222 paired planning and repeat CT scans (982 for training, 240 for testing). Compared with clinically used conventional affine or deformable registration methods, the proposed approach consistently demonstrated superior performance. The average registration time per image pair was reduced from 17.62 seconds to 0.22 seconds, corresponding to an approximately 80-fold speedup. Image alignment accuracy was substantially improved, with mean structural similarity (SSIM) increasing from 72.60% to 89.00%, corresponding to an improvement of approximately 22.6%. For clinically relevant target structures such as the clinical target volume (CTV), propagated contours achieved a Dice similarity coefficient of 58.20%, compared with 55.61% using conventional registration methods.
Conclusion
The proposed AI-driven DIR framework enables fast and accurate planning-to-repeat CT alignment across diverse proton therapy scenarios, substantially improving the efficiency and precision of routine clinical workflows.