DiffRecon: Diffusion-based CT reconstruction with cross-modal deformable fusion for DR-guided non-coplanar...

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Image fusion of intraoperative digital radiography (DR) and preoperative computed tomography (CT) is crucial for non-coplanar radiotherapy guidance.
• Current fusion methods struggle with unaligned data and dimensional discrepancies between DR and CT.
• CT reconstruction from DR could overcome these limitations.

Purpose of the Study:

• To develop a unified framework for intraoperative CT reconstruction from DR using a diffusion model.
• To improve the accuracy of image guidance in non-coplanar radiotherapy.
• To enhance the clinical applicability of DR-based CT reconstruction by incorporating preoperative CT as prior information.

Main Methods:

• Proposed DiffRecon, a unified generation and registration framework utilizing a diffusion model for CT reconstruction from DR.
• Employed a generation model for synthesizing intraoperative CTs and a registration model for aligning them.
• Designed a dual-encoder with cross-attention modules to learn from DR and preoperative CT, enabling 2D/3D feature conversion and cross-modal fusion.

Main Results:

• Achieved high image synthesis quality with RMSE of 0.02±0.01, PSNR of 44.92±3.26, and SSIM of 0.994±0.003.
• Demonstrated excellent dosimetric accuracy with mean gamma passing rates of 95.2% (1%/1 mm), 99.4% (2%/2 mm), and 99.9% (3%/3 mm).
• Validated accurate CT reconstruction from single DR projections.

Conclusions:

• DiffRecon accurately reconstructs CT from single DR projections with superior image generation and dosimetric accuracy.
• The method effectively addresses dimensional and alignment differences between DR and CT.
• DiffRecon shows significant potential for non-coplanar adaptive radiotherapy workflows.