PSCT-Net: Geometry-Aware Pediatric Skull CT Reconstruction via Differentiable Back-Projection and Attention-Guided Refinement

Researchers introduce PSCT-Net, a novel AI framework that reconstructs 3D pediatric skull CT scans from sparse 2D X-rays using differentiable back-projection and attention mechanisms, reducing radiation exposure to children while maintaining diagnostic accuracy. The team also releases PedSkull-CT, a new pediatric-focused dataset addressing the lack of child-specific medical imaging benchmarks in existing research.

PSCT-Net addresses a critical medical imaging challenge where reducing radiation exposure to pediatric patients conflicts with the need for accurate 3D diagnostic imaging. Traditional CT reconstruction from limited X-ray projections remains mathematically ill-posed, and prior approaches failed to properly model spatial relationships when lifting 2D features into 3D space, resulting in ambiguous depth estimates and degraded bone boundary definition. The framework's innovation lies in its geometry-aware approach: differentiable back-projection creates a spatially coherent volumetric prior that grounds 2D observations in 3D space, while the Attention-Guided Projection module learns complex non-linear mappings between image regions and volumetric coordinates. The Bidirectional Mamba module efficiently captures long-range volumetric dependencies without prohibitive computational costs.

This work matters significantly for pediatric radiology because it enables clinicians to diagnose craniofacial abnormalities—structural defects affecting facial and skull development—with substantially reduced radiation doses. Children face greater radiation sensitivity than adults, making dose reduction particularly crucial during critical developmental periods. The introduction of PedSkull-CT, a curated institutional cohort with both normal and pathological cases, fills an important research gap; most existing datasets focus on adult trunk imaging, leaving pediatric-specific models underdeveloped.

For AI researchers and medical device companies, this represents progress toward practical low-dose imaging pipelines that could be deployed clinically. The geometry-aware architecture may inspire similar approaches in other sparse-view 3D reconstruction problems beyond medical imaging. Future developments should focus on clinical validation and regulatory approval pathways, as healthcare adoption requires not just technical performance but safety certification and integration with existing hospital infrastructure.

• →PSCT-Net uses differentiable back-projection to create spatially accurate 3D skull reconstructions from limited 2D X-rays, reducing pediatric radiation exposure.
• →The framework combines attention-guided projection and efficient long-range volumetric modeling to improve bone boundary definition and depth disambiguation.
• →PedSkull-CT dataset addresses a critical research gap by providing pediatric-specific CT scans with both normal and pathological cases for model development.
• →Geometry-aware feature lifting represents a methodological advance over prior geometry-agnostic approaches that naively projected 2D information into 3D space.
• →Clinical translation potential exists for diagnosing pediatric craniofacial abnormalities with significantly lower radiation doses than conventional CT protocols.