Single Slice-to-3D Reconstruction in Medical Imaging and Natural Objects: A Comparative Benchmark with SAM 3D

A benchmark of five image-to-3D foundation models on medical and natural datasets showed that all methods suffer from low volumetric overlap due to severe depth ambiguity in single-slice inputs, though SAM3D best preserved topological similarity to ground-truth medical shapes. These findings demonstrate that reliable medical 3D reconstruction cannot rely on zero-shot transfer from natural-image priors and instead requires domain-specific adaptation and anatomical constraints. This project is built upon sam-3d-objects.

Installation

Follow the setup steps before running the following.

Datasets

You can download the datasets from their official websites, i.e., AeroPath, BTCV, Duke C-Spine, Medical Segmentation Decathlon, Google Scanned Objects, and Animal3D.

Preprocessing

For simplicity, we preprocess the data by isolating each object as an individual sample, e.g.,

cd notebook
python decompose_nii_to_obj_AeroPath.py

Benchmark 3D Reconstruction

Run the following commands

cd notebook
bash infer_and_eval.sh

Citing Our Work

If you find our work helpful for your research, please cite it using the following BibTeX entry.

@article{luo2026single,
  title={Single-Slice-to-3D Reconstruction in Medical Imaging and Natural Objects: A Comparative Benchmark with SAM 3D},
  author={Luo, Yan and Ravishankar, Advaith and Liu, Serena and Yang, Yutong and Wang, Mengyu},
  journal={arXiv preprint arXiv:2602.09407},
  year={2026}
}