🤖 AI

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

PSCT-Net is a geometry-aware deep learning framework that reconstructs 3D pediatric skull CTs from sparse bi-planar X-rays by integrating differentiable back-projection, attention-guided refinement, and a Bidirectional Mamba module to overcome depth ambiguity and improve bone boundary accuracy, while introducing a specialized pediatric dataset to address the lack of child-specific training data.

Original authors: Dong Yeong Kim, Jaewon Choi, Youmin Shin, Jungyu Lee, Myeongseop Kim, Jinwook Choi, Joo Whan Kim, Young-Gon Kim

Published 2026-06-19✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dong Yeong Kim, Jaewon Choi, Youmin Shin, Jungyu Lee, Myeongseop Kim, Jinwook Choi, Joo Whan Kim, Young-Gon Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Seeing the Whole Picture from Two Angles

Imagine you are trying to build a detailed 3D model of a child's skull, but you are only allowed to look at two flat, 2D X-ray photos: one from the front and one from the side.

This is a huge challenge for doctors. A standard CT scan gives a perfect 3D picture, but it uses a lot of radiation, which is dangerous for growing children. Doctors want to use the safer, low-radiation X-rays instead, but turning two flat photos into a 3D model is like trying to guess the shape of a hidden object just by looking at its shadow. It's a "guessing game" where the computer often gets lost, creating blurry shapes or inventing bones that don't exist (hallucinations).

The Solution: PSCT-Net

The authors created a new AI system called PSCT-Net. Think of it as a super-smart architect who doesn't just guess; they use the laws of physics to build the 3D skull correctly.

Here is how they did it, broken down into three main tricks:

1. The "Flashlight" Start (Differentiable Back-Projection)

Most old AI methods try to guess the 3D shape from scratch, which is like trying to sculpt a statue without a reference.

The Analogy: Imagine shining a flashlight through a flat photo of a face. The light rays travel straight through the photo into 3D space. Where the light hits, it leaves a "faint glow" that shows where the bone might be.
What they did: PSCT-Net uses a math trick called "differentiable back-projection" to trace these light rays. It creates a rough, blurry 3D "glow" of the skull first. This gives the AI a solid starting point that respects the actual physics of how X-rays work, so it doesn't get lost on where the front of the skull is versus the back.

2. The "Spotlight" Refinement (Attention-Guided Projection)

Once the AI has that rough 3D glow, it needs to sharpen the details. Old methods just copied the 2D picture onto the 3D model, which is like stamping a flat image onto a ball—it doesn't fit well.

The Analogy: Imagine a detective looking at a crime scene photo and asking, "If I see this specific spot on the 2D photo, where exactly does it belong in 3D space?"
What they did: They built a module called AGP-3D that acts like a smart spotlight. Instead of blindly copying pixels, it learns to connect specific parts of the 2D X-ray to the exact 3D location they belong to. This helps the AI figure out the tricky curves and thin bones that usually get blurred out.

3. The "Long-Range Memory" (Bidirectional Mamba)

To build a whole skull, the AI needs to understand how the left side connects to the right side, and the top connects to the bottom. Standard AI models often get "tunnel vision" and forget the big picture, or they get so slow they can't handle the whole image at once.

The Analogy: Imagine trying to remember a long story. Some people can only remember the last sentence they heard. Others can remember the whole story from start to finish instantly.
What they did: They used a new type of AI engine called BiM-3D (based on "Mamba"). It's like a reader who can scan the entire 3D skull from every angle at once, remembering how every bone relates to every other bone, but doing it super fast (like reading a book in seconds instead of hours).

The New "Training School" (PedSkull-CT)

The paper also mentions a major problem with previous AI training: most AI was trained on adult bodies (like spines or chests), which look very different from a baby's skull. Babies have soft spots (fontanelles) and thinner bones.

The Analogy: It's like trying to teach someone to drive a Formula 1 car by only letting them practice on a tricycle. The skills don't transfer well.
What they did: The team created a brand new, private dataset called PedSkull-CT. It contains 982 real scans of children's skulls (both healthy and with medical issues). They used this to "train" their AI specifically on how baby skulls look, ensuring it learns the right details.

The Results

When they tested PSCT-Net:

It worked better than any other method on public tests (for lungs, spines, and pelvises).
On their new baby skull dataset, it produced the clearest, most accurate 3D images, preserving tiny details that other methods missed.
It was fast enough to be useful in a real hospital, unlike some newer methods that take too long to compute.

Summary

In short, the authors built a tool that turns two low-radiation X-rays into a high-quality 3D skull model for children. They did this by:

Using physics to create a rough 3D guide first.
Using smart attention to connect 2D dots to 3D spots accurately.
Using a fast memory system to understand the whole skull at once.
Training it on a special collection of real baby skull scans.

This offers a way to see inside a child's head clearly without exposing them to dangerous amounts of radiation.

Technical Summary: PSCT-Net for Pediatric Skull CT Reconstruction

Problem Statement

Computed Tomography (CT) is the gold standard for diagnosing pediatric craniofacial abnormalities, yet the associated ionizing radiation poses severe risks to developing anatomies due to heightened radiosensitivity and long life expectancy for potential malignancies. While bi-planar X-ray imaging offers a low-dose alternative, it lacks the volumetric depth information necessary to assess complex skull deformities. Reconstructing high-fidelity 3D CT volumes from sparse 2D X-ray projections is a severely ill-posed inverse problem.

Existing deep learning approaches typically rely on geometry-agnostic feature lifting, where 2D features are naively replicated or linearly projected into 3D space. These implicit mechanisms fail to model physical acquisition geometry, leading to spatial misalignment, depth ambiguity, and the loss of fine-grained osseous structures (e.g., sutures and fontanelles) critical for pediatric diagnosis. Furthermore, recent diffusion-based models, while improving textural realism, impose prohibitive computational costs due to iterative denoising, rendering them impractical for time-sensitive clinical workflows. Additionally, a significant domain gap exists: public benchmarks focus on adult trunk anatomy (lung, spine, pelvis) and lack pediatric-specific physiological markers such as unclosed fontanelles and thinner cortical bone.

Methodology: PSCT-Net

The authors propose PSCT-Net, a geometry-aware framework that integrates explicit geometric priors with computationally efficient context modeling. The architecture is built upon a standard 2D-to-3D conditional Generative Adversarial Network (cGAN) but introduces four key innovations to enforce spatial consistency and robust global context:

1. Differentiable Back-Projection Initialization

To mitigate depth ambiguity, the network initializes a coarse volumetric prior via a differentiable back-projection layer. This layer traces X-ray intensities along physical ray paths defined by the projection matrix (rotation and translation), generating a geometrically faithful attenuation volume ( $V_{prior}$ ). This step explicitly injects acquisition geometry into the network, providing a spatially faithful initialization that alleviates depth ambiguity from the outset.

2. Geometry-Aware Multi-view Conditioning

The framework employs a dual conditioning strategy to enforce geometric consistency at both encoding and decoding stages:

BP-C Module (Encoder): Back-projects 2D feature maps into 3D space according to acquisition geometry. These view-specific volumes are averaged to form a global prior, which is concatenated with main encoder features to resolve depth ambiguity early.
MV3D-C Module (Decoder): Aligns high-level volumetric features from view-specific branches into a common coordinate system. These aligned features are averaged and concatenated with the main decoder to ensure the final reconstruction remains semantically consistent with input projections.

3. Attention-Guided Projection (AGP-3D)

Replacing fixed linear projections or simple feature replication, the AGP-3D module leverages a Multi-Head Attention (MHA) mechanism to learn non-linear voxel-wise correspondences between 2D image regions and 3D spatial locations. By treating 3D grid locations as queries and 2D features as keys, the network dynamically learns where to retrieve relevant texture information, enabling discriminative feature aggregation rather than blind projection.

4. Bidirectional Mamba (BiM-3D)

To capture holistic cranial geometry efficiently, the framework incorporates a Bidirectional Mamba (BiM-3D) module. This module utilizes a Bidirectional Selective State Space Model (Bi-SSM) to model long-range volumetric dependencies with linear complexity ( $O(N)$ ), avoiding the quadratic computational cost ( $O(N^2)$ ) of standard Transformers while overcoming the limited receptive fields of convolutions.

Training Objective

The network is trained using a compound loss function comprising:

Adversarial Loss ( $L_{adv}$ ): A conditional Least Squares GAN with a 3D PatchDiscriminator.
Reconstruction Loss ( $L_{rec}$ ): $\ell_1$ voxel-wise reconstruction loss.
Projection Consistency Loss ( $L_{proj}$ ): Enforces consistency via 2D orthogonal projections.

Key Contributions

Novel Framework: PSCT-Net is the first framework to integrate differentiable back-projection with state-space modeling for X-ray to CT reconstruction. By explicitly encoding acquisition geometry, it resolves the depth ambiguity that causes existing approaches to hallucinate incorrect structures.
PedSkull-CT Dataset: The authors curated PedSkull-CT, a private institutional pediatric skull CT cohort comprising 982 scans (ages 1–24 months) covering both normal and pathological cases with paired simulated X-rays. This addresses the lack of pediatric-specific datasets in existing public benchmarks.
Performance and Efficiency: The method achieves state-of-the-art performance across three public benchmarks and the private pediatric cohort. Crucially, it outperforms diffusion-based methods while maintaining the high inference efficiency inherent to single-step architectures, making it suitable for clinical deployment.

Experimental Results

The authors evaluated PSCT-Net on three public benchmarks (LIDC-IDRI, CTSpine1K, CTPelvic1K) and the private PedSkull-CT cohort.

Public Benchmarks: On LIDC-IDRI, PSCT-Net achieved a PSNR of 27.18 dB, surpassing the diffusion-based DiffuX2CT by 0.83 dB. On CTPelvic1K, it achieved 33.06 dB, outperforming the second-best method by 1.35 dB. These results demonstrate robust generalization across diverse anatomical regions.
PedSkull-CT: On the private pediatric cohort, PSCT-Net outperformed all baselines, improving PSNR by 1.28 dB and SSIM by 0.022 over the second-best method (X2CT-GAN).
Ablation Studies: Experiments confirmed that each module contributes to performance. Notably, the BiM-3D module alone provided the largest gain (+1.04 dB), highlighting the importance of global context modeling. The full model achieved the best results, confirming the complementarity of geometric priors and long-range modeling.
Real-World Generalization: The model successfully reconstructed volumes from real-world clinical X-rays unseen during training, preserving patient-specific features such as mandible curvature and orbital socket depth.

Significance and Limitations

The paper posits that geometry-aware reconstruction offers a viable path toward low-dose pediatric imaging. By explicitly modeling acquisition geometry, PSCT-Net avoids the "black-box" transformation of 2D-to-3D lifting, thereby reducing anatomical hallucinations and preserving critical fine-grained structures.

The authors acknowledge limitations: while the method recovers global geometry and major osseous structures with high fidelity, resolving sub-millimeter details (such as fine cranial sutures) remains challenging due to fixed voxel resolution. Future work is proposed to incorporate patch-based refinement and Implicit Neural Representations or 3D Gaussian primitives to recover finer details, alongside validation through reader studies on craniosynostosis diagnosis.