DuFal: Dual-Frequency-Aware Learning for High-Fidelity Extremely Sparse-View CBCT Reconstruction

Cuong Tran^1*     Trong-Thang Pham^2*     Ngoc-Son Nguyen¹     Duy Minh Ho Nguyen³     Ngan Le^2†

¹ FPT Software AI Center, Vietnam    ² AICV Lab, EECS Department, University of Arkansas    ³ IMPRS-IS, University of Stuttgart, and DFKI, Germany

^* Equal contribution    ^† Corresponding author

---

Overview

Sparse-view Cone-Beam Computed Tomography (CBCT) reconstruction remains challenging due to severe undersampling of high-frequency anatomical details. Conventional CNN-based methods are often biased toward low-frequency information, leading to the loss of fine structures.

We propose DuFal (Dual-Frequency-Aware Learning), a novel dual-path framework that jointly exploits spatial- and frequency-domain representations. At its core is a High-Local Factorized Fourier Neural Operator, which consists of:

• a global high-frequency enhanced branch for capturing long-range spectral patterns, and
• a local high-frequency enhanced branch that processes spatial patches to preserve locality.

To improve efficiency, DuFal introduces spectral–channel factorization to reduce model complexity, along with a cross-attention frequency fusion module for effective integration of spatial and spectral features. The fused representations are decoded into projection features and reconstructed into 3D volumes via an intensity field decoding pipeline.

Experiments on LUNA16 and ToothFairy demonstrate that DuFal consistently outperforms state-of-the-art methods, particularly under extremely sparse-view conditions, with superior preservation of high-frequency anatomical structures.

*Overview of DuFal architecture*

🏆 Key Contributions

• Dual-path encoding with spatial and frequency branches for high-frequency detail preservation.
• HiLocFFNO blocks combine global and local frequency modeling in a modular Frequency Encoder.
• Spectral-Channel Factorization (SCF) reduces FNO parameters while retaining quality.
• Cross-Attention Frequency Fusion (CAFF) merges spatial and spectral features in the frequency domain.
• Evaluated on LUNA16 and ToothFairy in extremely sparse-view settings.

---

Table of Contents

• 📣 News
• 🛠️ Requirements and Installation
• 📄 Configuration
• 🧩 Creating splits for training and testing
• 🧪 Usage
• 🙏 Acknowledgments
• 📖 Citation
• ⚠️ Usage and License Notices

---

📣 News

• [Dec 2025]: DuFal paper has been accepted with minor revisions at TMLR!

---

🛠️ Requirements and Installation

pip install torch==1.13 pytorch3d SimpleITK easydict

---

📄 Configuration

Edit configs/config.yaml to set your dataset path (after preprocessing):

dataset:
  root_dir: /path/to/your/data

---

---

🧩 Creating splits for training and testing

Dataset preparation and split definitions are documented in data/README.md. Use the official splits for LUNA16 and ToothFairy (see data/LUNA16/ and data/ToothFairy/).

---

🧪 Usage

Training (single GPU)

CUDA_VISIBLE_DEVICES=0 python code/train.py \
  --batch_size 1 \
  --epoch 600 \
  --dst_name ToothFairy \
  --num_views 10 \
  --random_view \
  --cfg_path ./configs/config.yaml \
  --num_workers 8 \
  --eval_interval 50 \
  --save_interval 50 \
  --setting spatial-test \
  -trunc LL-LH LL-LH LL-LH LL-LH \
  -sobel 0 0 0 0 \
  -patch 1 1 1 1 \
  -psize 16 16 16 16 \
  -fac none none dep-sep dep-sep \
  -attn 0 0 0 1 \
  -prop 0 0 0 0 \
  -swin 0 \
  -wsize 0 0 0 0 \
  -grid linear \
  -fuse 8 \
  -fno \
  -skip add \
  --use_wandb

Key options:

• --dst_name: dataset name (LUNA16 or ToothFairy).
• --num_views: number of projection views.
• --cfg_path: path to the config file.
• --setting: run name used for logging and outputs.
• -trunc/-sobel/-patch/-psize/-fac/-attn/-prop/-swin/-wsize/-grid/-fuse/-fno/-skip: architecture and frequency-encoding settings (kept consistent with the paper config).

Testing

CUDA_VISIBLE_DEVICES=1 python code/evaluate.py --epoch xxx --dst_name {LUNA16 or ToothFairy} --split test --num_views 10 --out_res_scale 1.0 --setting xxx

---

🙏 Acknowledgments

This project builds upon the DIF-Gaussian codebase.

---

📖 Citation

If you find this work useful, please cite our paper:

@article{van2026dufal,
  title={DuFal: Dual-Frequency-Aware Learning for High-Fidelity Extremely Sparse-view {CBCT} Reconstruction},
  author={Cuong Tran Van and Trong-Thang Pham and Ngoc-Son Nguyen and Duy Minh Ho Nguyen and Ngan Le},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2026},
  url={https://openreview.net/forum?id=2wAZjAtK16},
  note={J2C Certification}
}

---

⚠️ Usage and License Notices

The model is not intended for clinical use as a medical device, diagnostic tool, or any technology for disease diagnosis, treatment, or prevention. It is not a substitute for professional medical advice, diagnosis, or treatment. Users are responsible for evaluating and validating the model to ensure it meets their needs before any clinical application.