X2Shape: CT-free 3D multi-organ reconstruction with biplanar X-rays.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Reconstructing 3D anatomy from 2D X-rays is challenging but offers lower radiation exposure than CT.
• Existing methods often rely on CT data or lack accuracy and robustness.

Purpose of the Study:

• To develop a deep learning framework (X2Shape) for direct 3D multi-organ reconstruction from biplanar X-rays.
• To overcome data scarcity using a novel augmentation strategy.
• To achieve accurate and robust 3D reconstruction without CT priors.

Main Methods:

• Developed X2Shape, a deep learning framework utilizing geometry-aware volumetric backprojection and a state-space model-based cross-view fusion module.
• Implemented a hybrid deformation-based augmentation strategy for generating diverse training data.
• Validated the framework on two thoracic benchmarks (TotalSegmentator-Subset and LCTSC).

Main Results:

• X2Shape achieved high accuracy, with Dice scores of 88.98% and 75.62% on the tested datasets.
• Demonstrated substantial improvement over existing methods.
• Showcased strong cross-dataset generalization and efficient, robust reconstruction of diverse organ structures.

Conclusions:

• X2Shape provides accurate 3D organ reconstruction from X-rays, eliminating the need for CT.
• This scalable paradigm offers a low-cost, low-radiation solution for 3D imaging.
• Potential applications include personalized diagnostics, surgical planning, and image-guided interventions.