Add model card

README.md

@@ -0,0 +1,84 @@
+---
+library_name: pytorch
+tags:
+  - medical-imaging
+  - computed-tomography
+  - ct-reconstruction
+  - sparse-view-ct
+  - compressed-sensing
+  - neural-operator
+  - inverse-problems
+pipeline_tag: image-to-image
+---
+
+<!-- TODO: add arXiv badge once the preprint is public, e.g.
+[![arXiv](https://img.shields.io/badge/arXiv-XXXX.XXXXX-b31b1b.svg?style=flat-square&logo=arxiv)](https://arxiv.org/abs/XXXX.XXXXX) -->
+[![Code](https://img.shields.io/badge/Code-neuraloperator%2Fsparse__ct-black?style=flat-square&logo=github)](https://github.com/neuraloperator/sparse_ct)
+
+# Resolution-Agnostic Neural Operators for Multi-Rate Sparse-View CT
+
+> **Resolution-Agnostic Neural Operators for Multi-Rate Sparse-View CT**
+> Aujasvit Datta\*, Jiayun Wang\*, Asad Aali, Anima Anandkumar
+> *Caltech · Georgia Tech · IIT Kanpur · Stanford University* — \*equal contribution
+
+**CTO (Computed Tomography neural Operator)** is the first neural operator (NO) framework for CT reconstruction. Sparse-view CT (SVCT) reconstructs an image from a limited number of X-ray projections to reduce radiation dose and scan time — an ill-posed inverse problem. Existing learning methods (e.g. CNNs) overfit to a fixed acquisition setup and fail to generalize across sampling rates and resolutions. CTO instead learns the reconstruction map in **continuous function space**, so a **single model generalizes across measurement sampling rates and image resolutions without retraining**.
+
+## Highlights
+
+- **First neural operator for CT reconstruction** — moves from fixed discretized grids to a continuous function-space formulation.
+- **Multi-rate with one model** — a single CTO handles a range of sparse-view acceleration rates (**10×, 20×, 40×, 80×**, i.e. ≈72/36/18/9 of 720 projection views) instead of one model per rate.
+- **Resolution-agnostic** — built on DISCO (Discrete–Continuous) convolutions whose kernels live in function space, giving a consistent receptive field across resolutions (discretization-convergent).
+- **Dual-domain reconstruction** — complementary neural operators in the **sinogram** domain (polar coordinates, with joint spatial + frequency branches) and the **image** domain (Cartesian grid), coupled by data-consistency/physical-update terms in an unrolled network.
+- **Rotation-equivariant** — learning in the sinogram's polar coordinates makes object rotation correspond to a shift along the angular axis, improving robustness to acquisition orientation.
+- **Strong results** — >3.4 dB PSNR gain over CNN baselines in multi-resolution settings; vs. state-of-the-art diffusion methods, **~500× faster inference** with an average **+3 dB**. Also shows out-of-distribution robustness under cross-dataset transfer and noisy sinograms.
+
+## Models
+
+This repository provides two pretrained CTO checkpoints (PyTorch Lightning), one per dataset:
+
+| Checkpoint | Dataset | Forward model | Config |
+|---|---|---|---|
+| [`aapm/cto_aapm.ckpt`](./aapm/cto_aapm.ckpt) | [AAPM Low-Dose CT Grand Challenge](https://www.aapm.org/grandchallenge/lowdosect/) | `AAPMRadonTransform` (720 projections) | [`aapm/cto_aapm_config.yaml`](./aapm/cto_aapm_config.yaml) |
+| [`kits/cto_kits.ckpt`](./kits/cto_kits.ckpt) | [C4KC-KiTS (kidney CT)](https://kits-challenge.org/) | `KitsRadonTransform` (720 projections) | [`kits/cto_kits_config.yaml`](./kits/cto_kits_config.yaml) |
+
+Both checkpoints share the same architecture and were trained/validated at **256×256** resolution with an MSE objective (up to 100 epochs), and cover training acceleration rates `[10, 20, 40, 80]`.
+
+## Architecture
+
+CTO is an end-to-end **unrolled network** operating in function space (see paper, Fig. 2):
+
+1. **Sinogram operator (NO\_s)** — `SinoCircularUDNORFFTReconstructor`, a U-shaped DISCO block (UDNO) in polar coordinates `(r, θ)` that jointly processes the sub-sampled sinogram in the **spatial** domain and the **frequency** domain (1D FFT along the detector axis `r`). This unifies different angular sampling rates.
+2. **Inverse Radon (R⁻¹)** — maps the completed sinogram back to image space.
+3. **Image operator (NO\_i) + data consistency** — `ImageNOVarnetReconstructor`, an unrolled variational network (**3 cascades**) that refines the image with an image-space DISCO operator and a physical data-consistency update against the measurements.
+
+Key hyperparameters (from the configs): DISCO kernels use a **piecewise-linear basis**, `chans=32`, `num_pool_layers=4`, `kernel_shape=[6, 7]`, `radius_cutoff=0.02`; the sampler is `VarnetSparseCTSampler`; `num_projections=720`.
+
+## Usage
+
+Download the weights and configs with the Hugging Face CLI or `huggingface_hub`:
+
+```bash
+hf download aujasvit/cto --local-dir ./cto
+```
+
+```python
+from huggingface_hub import hf_hub_download
+
+ckpt = hf_hub_download("aujasvit/cto", "aapm/cto_aapm.ckpt")   # or "kits/cto_kits.ckpt"
+cfg  = hf_hub_download("aujasvit/cto", "aapm/cto_aapm_config.yaml")
+```
+
+The checkpoints are PyTorch Lightning `CTLightningModule` states; loading them into the model and running inference requires the model code and scripts from the project repository:
+
+**Code:** https://github.com/neuraloperator/sparse_ct
+
+## Citation
+
+<!-- TODO: update venue / add arXiv id once available -->
+```bibtex
+@article{datta2026cto,
+  title  = {Resolution-Agnostic Neural Operators for Multi-Rate Sparse-View CT},
+  author = {Datta, Aujasvit and Wang, Jiayun and Aali, Asad and Anandkumar, Anima},
+  year   = {2026}
+}
+```