Resolution-Agnostic Neural Operators for Multi-Rate Sparse-View CT
_{ECCV 2026}

Resolution-Agnostic Neural Operators for Multi-Rate Sparse-View CT
ECCV 2026
Aujasvit Datta*, Jiayun Wang*, Asad Aali, Anima Anandkumar
arXiv: https://arxiv.org/abs/2512.12236

All common tasks are wrapped as make targets. Run make help to list them. Every target is a thin wrapper around a python ... command if you prefer to run things directly.

1. Configuration

Before running anything, review configs/base.yaml and update it for your setup. Common changes:

• use_wandb: set to false to disable Weights & Biases logging for both train and test.
• logger.project and logger.entity: your Weights & Biases project and account. Leave entity as null to use your logged-in default, or set WANDB_ENTITY.
• trainer.num_devices: the number of GPUs to use.
• trainer.max_epochs and trainer.lr: training hyperparameters.

Dataset-specific settings (model, sampler, batch size, data paths) live in configs/aapm.yaml and configs/kits.yaml.

2. Install

conda create --name cto_env python=3.11
conda activate cto_env
make setup          # install python dependencies
make torch-radon    # build torch-radon (needs a CUDA build environment: nvcc + toolkit)

Tested with Python 3.11 on Linux. make torch-radon clones torch-radon, applies the required patch from torch-radon_fix/, and builds it.

3. Data

The Low-dose CT AAPM dataset is substantially smaller than the C4KC-KiTS dataset, hence easier to work with. Download the raw data first, then preprocess it:

• AAPM: download from https://aapm.app.box.com/s/eaw4jddb53keg1bptavvvd1sf4x3pe9h/folder/144226105715, then unzip FD_1mm.zip to full_1mm/.
• C4KC-KiTS: download from https://www.cancerimagingarchive.net/collection/c4kc-kits/.

To preprocess the data, run:

make data-aapm AAPM_RAW_DIR=/path/to/full_1mm
make data-kits KITS_RAW_DIR=/path/to/C4KC-KiTS

Preprocessed data is written to data/aapm/ and data/kits/ (matching the configs).

4. Test the pretrained models

Download our pretrained weights and test them (needs the corresponding preprocessed dataset from step 2):

make weights        # download pretrained weights from Hugging Face -> ./weights
make test-aapm      # test the AAPM model
make test-kits      # test the C4KC-KiTS model

Metrics are printed and saved to results/ as CSV + JSON. No Weights & Biases account is required.

5. Train a model

make train-aapm
make train-kits

Trained models are saved under models/. Training logs to Weights & Biases by default; set WANDB_ENTITY=<your-username> to choose the entity, or set use_wandb: false in configs/base.yaml to turn logging off.

To test a model you trained yourself instead of the pretrained weights, run python scripts/test.py -c configs/aapm.yaml without --fix init_exp_dir, so it reads the checkpoint from the config's exp_dir under models/.