Neural Discrete Representation Learning for Sparse-View CBCT Reconstruction: From Algorithm Design to Prospective Multicenter Clinical Evaluation

Haoshen Wang, Lei Chen, Wei-Hua Zhang, Linxia Wu, Yong Luo, Zengmao Wang, Yuan Xiong, Chengcheng Zhu, Wenjuan Tang, Xueyi Zhang, Wei Zhou, Xuhua Duan, Lefei Zhang, Gao-Jun Teng, Bo Du, Huangxuan ZhaoarXiv 2025 · DOI · arXiv

Introduction

Cone Beam Computed Tomography (CBCT) is a cornerstone of modern interventional radiology, particularly for guiding percutaneous thoracic punctures such as lung biopsies and tumor ablations. However, the high radiation dose associated with standard CBCT scans remains a significant concern, potentially increasing the risk of secondary cancers in patients. While reducing the number of X-ray projections (sparse-view sampling) can lower radiation exposure, it typically results in severe image degradation. Standard reconstruction algorithms, such as the Feldkamp-Davis-Kress (FDK) method, produce images marred by streak artifacts and motion blur when data is missing.

Existing deep learning solutions often attempt to map low-dose inputs directly to high-dose outputs. However, these methods frequently suffer from "hallucinations"—the generation of anatomically incorrect structures—and lack the rigorous, multicenter prospective validation necessary for clinical deployment. This paper introduces DeepPriorCBCT, a framework designed to reconstruct high-quality images from just one-sixth of the standard radiation dose by leveraging learned anatomical priors.

Methodology: DeepPriorCBCT

DeepPriorCBCT shifts the paradigm from generative interpolation to anatomical retrieval. The framework utilizes a three-stage process based on neural discrete representation learning:

  1. High-Dose Prior Learning: The system utilizes an autoencoder with a discrete codebook (similar to VQ-VAE architectures) trained on high-quality, full-view CBCT scans. This process compresses continuous anatomical features into discrete vectors, creating a "dictionary" of healthy, high-resolution anatomical structures.
  2. Low-Dose Reconstruction: When presented with a sparse-view (low-dose) input, the network does not attempt to generate missing pixels from scratch. Instead, a code classifier maps the low-quality features to their closest high-quality counterparts within the pre-trained codebook.
  3. Feature Fusion: Multi-resolution features from the low-dose input are fused with the high-quality features retrieved from the codebook. This ensures the final reconstruction maintains the specific global structure of the patient while filling in fine-grained details using physically accurate priors.

By using retrieval over hallucination, DeepPriorCBCT identifies anatomical structures and pulls clean versions from its memory, avoiding the smoothing of small tumors or the creation of artificial structures common in end-to-end CNNs.

Experimental Design and Multicenter Validation

The study's validation is uniquely robust, utilizing both retrospective and prospective cohorts:

  • Retrospective Cohort: Data from 4,102 patients (8,675 CBCT scans) were collected from 12 medical centers. This was used for training (2,987 patients), validation (100 patients), and testing (1,015 patients from 4 unseen centers) to ensure the model's generalizability across different hospital environments.
  • Prospective Cohort (COOS Trial): A prospective clinical trial involved 138 patients across 5 centers scheduled for thoracic punctures. During their procedures, each patient received both a standard full-view scan and a 1/6-view undersampled scan to directly compare the AI's performance in a live surgical setting.

Results

Quantitative and Clinical Performance

DeepPriorCBCT demonstrated the ability to operate effectively at 1/6th of the standard views. This resulted in a dramatic reduction in the Dose Area Product (DAP), falling from 1674.24 μGym² (standard) to 299.47 μGym². Quantitatively, the model achieved a PSNR of ~28.96 and an SSIM of ~0.79, significantly outperforming standard clinical methods and baseline AI models like RED-CNN and CycleGAN.

The Visual Turing Test and Surgical Utility

To bridge the gap between mathematical metrics and clinical reality, the authors conducted a "Visual Turing Test." Eleven physicians (radiologists and interventionalists) were asked to distinguish between actual full-dose images and DeepPriorCBCT’s 1/6-dose reconstructions. The physicians' accuracy was near random chance (Kappa < 0.1), indicating the images were indistinguishable.

In the prospective trial, 25 interventionalists evaluated the images for surgical guidance. The preference was split nearly 50/50 between the full-dose standard and the 1/6-dose AI reconstruction, successfully proving the non-inferiority of the low-dose AI approach in real-world clinical workflows.

Discussion and Limitations

While DeepPriorCBCT represents a significant leap forward in low-dose imaging, the study notes several areas for future work. The current validation was strictly limited to thoracic lesions; further testing is required to confirm the model's efficacy in other regions, such as the head or abdomen. Additionally, while the 138-patient prospective trial is substantial, larger cohorts are needed to establish long-term clinical benefits regarding radiation-induced risks. Finally, the authors plan to conduct Randomized Controlled Trials (RCTs) to determine if the use of DeepPriorCBCT improves surgical efficiency, such as reducing procedure time or complication rates.

Summary updated

Abstract

Cone beam computed tomography (CBCT)-guided puncture has become an established approach for diagnosing and treating early- to mid-stage thoracic tumours, yet the associated radiation exposure substantially elevates the risk of secondary malignancies. Although multiple low-dose CBCT strategies have been introduced, none have undergone validation using large-scale multicenter retrospective datasets, and prospective clinical evaluation remains lacking. Here, we propose DeepPriorCBCT - a three-stage deep learning framework that achieves diagnostic-grade reconstruction using only one-sixth of the conventional radiation dose. 4102 patients with 8675 CBCT scans from 12 centers were included to develop and validate DeepPriorCBCT. Additionally, a prospective cross-over trial (Registry number: NCT07035977) which recruited 138 patients scheduled for percutaneous thoracic puncture was conducted to assess the model's clinical applicability. Assessment by 11 physicians confirmed that reconstructed images were indistinguishable from original scans. Moreover, diagnostic performance and overall image quality were comparable to those generated by standard reconstruction algorithms. In the prospective trial, five radiologists reported no significant differences in image quality or lesion assessment between DeepPriorCBCT and the clinical standard (all P>0.05). Likewise, 25 interventionalists expressed no preference between model-based and full-sampling images for surgical guidance (Kappa<0.2). Radiation exposure with DeepPriorCBCT was reduced to approximately one-sixth of that with the conventional approach, and collectively, the findings confirm that it enables high-quality CBCT reconstruction under sparse sampling conditions while markedly decreasing intraoperative radiation risk.
BibTeX
@misc{wang2025neural,
  author = {Haoshen Wang and Lei Chen and Wei-Hua Zhang and Linxia Wu and Yong Luo and Zengmao Wang and Yuan Xiong and Chengcheng Zhu and Wenjuan Tang and Xueyi Zhang and Wei Zhou and Xuhua Duan and Lefei Zhang and Gao-Jun Teng and Bo Du and Huangxuan Zhao},
  title = {{Neural Discrete Representation Learning for Sparse-View CBCT Reconstruction: From Algorithm Design to Prospective Multicenter Clinical Evaluation}},
  year = {2025},
  eprint = {2512.00873},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.00873},
}
Figure