DeepSparse

Cone-beam computed tomography is the workhorse 3D imaging modality of dentistry, orthopedics, and image-guided radiotherapy, but a standard scan fires hundreds of X-ray projections at the patient. Sparse-view reconstruction cuts that dose by an order of magnitude — six to ten projections instead — and relies on a model to recover the missing structure. The catch has been generalization: methods that optimize a neural field per scan are slow, and supervised networks trained on one anatomy and scanner geometry degrade badly when moved to another. Every new site has effectively meant a new model and a new dataset.

DeepSparse, from the xmed-lab group led by Xiaomeng Li at the Hong Kong University of Science and Technology with a collaborator at Sichuan University, is presented as the first foundation model for sparse-view CBCT reconstruction. It transplants the pretrain-once, adapt-cheaply pattern into tomographic reconstruction: a single backbone is pretrained on thousands of CT volumes covering the body from head to knee, then fine-tuned in two short steps onto a target anatomy and scanner geometry using a small number of labeled scans. The paper was accepted to IEEE Transactions on Medical Imaging in 2026.

Two components carry the result. DiCE (Dual-Dimensional Cross-Scale Embedding) is the reconstruction network, fusing multi-view 2D features from the projections with multi-scale 3D features back-projected into the volume. HyViP (Hybrid View Sampling Pretraining) randomizes the number of input views during pretraining, so the backbone is not locked to one acquisition protocol.

#Key Features

• Anatomy-agnostic pretraining: A single backbone is pretrained across head, chest, abdomen, pelvis, and knee anatomy drawn from 26 hospitals worldwide, then adapted to specific targets rather than trained from scratch per site.
• Variable view counts by construction: HyViP samples between 6 and 24 projections during pretraining, generating 2D features from a random subset while computing 3D features from the maximum, so the model tolerates whichever sparse protocol a clinic uses.
• Two-step fine-tuning: Step one adapts the pretrained backbone to the target dataset; step two adds denoising layers that align sparse-view 3D features with dense-view reference features under an L1 loss, refining detail the first step cannot recover.
• Small and fast: At 7.2M parameters DeepSparse uses roughly one seventh the parameters of the previous state of the art and reconstructs 7.6 times faster, finishing a volume in seconds.
• Codebook-based implicit representation: Vector quantization in the encoder-decoder lets the model reconstruct volumes at arbitrary resolution while keeping inference memory bounded.

#Technical Details

DiCE pairs a 2D encoder that extracts multi-scale features from the input projections with a 3D decoder that produces a low-resolution volumetric representation from back-projected features, plus a point decoder that predicts attenuation coefficients at arbitrary coordinates — an implicit neural representation, which is why reconstruction resolution is not fixed by the architecture. The published configuration uses four feature scales and a volumetric resolution of 32. Pretraining ran on AbdomenAtlas-8K: 5,195 CT scans cropped along the axial direction into 8,407 sub-volumes, resampled to 256³ at 1.5 mm isotropic spacing, with digitally reconstructed radiographs generated under a standard cone-beam geometry.

Fine-tuning and evaluation covered five target datasets across four anatomies — LUNA16 (888 chest CT), ToothFairy (443 head CBCT), PANORAMA (2,044 abdominal CT), PENGWIN (100 pelvic CT), and a 614-scan knee CBCT set. At six views on LUNA16, DeepSparse reaches 30.22 dB PSNR and 89.96 SSIM against 29.23 dB and 87.47 for C²RV, the prior best. Gains run 1–2 dB PSNR on head, chest, and abdomen and 3–4 dB on pelvis and knee. The value of pretraining shows most clearly under data scarcity: with only 73 training scans, 10% of LUNA16, the pretrained model holds 28.35 dB versus 27.13 dB for the same architecture trained from scratch.

#Applications

The direct beneficiary is any imaging workflow where dose matters and scan volume is high — pediatric and dental CBCT, intraoperative and interventional imaging, and daily cone-beam positioning scans in radiotherapy, where patients are imaged repeatedly over a treatment course. Because adaptation needs only a two-step fine-tune on a modest labeled set, a department can bring the model onto its own scanner geometry without assembling a large in-house corpus — the usual obstacle for deep reconstruction methods outside well-resourced centers.

#Impact

DeepSparse extends the foundation-model framing from medical image interpretation, where it is now routine, into image formation, where per-dataset training from scratch had remained the norm. Code for pretraining, fine-tuning, and evaluation is released under an MIT license, and both pretrained and fine-tuned checkpoints are published on Hugging Face, making it directly reusable rather than a paper-only result. The evaluation is entirely retrospective and simulation-based: projections are digitally reconstructed from CT volumes rather than acquired on a physical cone-beam scanner, so real-world scatter, motion, and detector noise remain untested, and no reader study or clinical validation has been reported.