---
license: cc-by-nc-sa-4.0
datasets:
- ibrahimhamamci/CT-RATE
tags:
- chest-ct
- radiology
- 3d-medical-imaging
- medical
- ct-rate
- abnormality-classification
- computer-vision
---
Structured Spectral Graph Representation Learning for Multi-label Abnormality Analysis from 3D CT Scans ๐ฉบ๐จ๐ปโโ๏ธ
โ
CT-SSG model weights, trained on CT-RATE.
๐ Accepted at MELBA 2026: [https://arxiv.org/abs/2510.10779](https://arxiv.org/abs/2510.10779).
โก๏ธ PyTorch implementation available at [https://github.com/theodpzz/ct-ssg](https://github.com/theodpzz/ct-ssg).
## ๐ฅ Available resources
**model_state_dict.pt**: Model weights for CT-SSG trained on the **CT-RATE training set**.
**thresholds.json**: Per-abnormality classification thresholds optimized on **our internal CT-RATE validation set**. The **official CT-RATE test set** was not used during threshold optimization to preserve unbiased evaluation.
## โ ๏ธ Splits
Since the CT-RATE dataset does not provide an official train/validation/test split, we adopt the following protocol for training and evaluation of the released model.
The test set remains strictly untouched and corresponds to the official validation partition of CT-RATE, i.e., all patients whose identifiers contain the tag 'valid'.
For internal validation, we use a subset of the original training data consisting of patients with IDs ranging from 1 to 1308 (denoted as train_1 to train_1308).
The released model is trained on the remaining portion of the CT-RATE training set, excluding the aforementioned validation subset. In other words, all training samples outside the ID range 1-1308 are used for model training.
## ๐ค๐ป Acknowledgment
We thank contributors from the CT-RATE dataset available at [https://huggingface.co/datasets/ibrahimhamamci/CT-RATE](https://huggingface.co/datasets/ibrahimhamamci/CT-RATE), from the Rad-ChestCT dataset available at [https://zenodo.org/records/6406114](https://zenodo.org/records/6406114) and from the Merlin Abdominal CT dataset available at [https://stanfordaimi.azurewebsites.net/categories/datasets?domain=BODY](https://stanfordaimi.azurewebsites.net/categories/datasets?domain=BODY).
## Purpose
The model, trained on a publicly available dataset, is provided for academic and research purposes only, to support reproducibility of the results described in the associated paper. This repository is a research prototype, and is not intended for clinical use.
## ๐Citation
If you find this repository useful for your work, we would appreciate the following citation:
```bibtex
@article{dipiazza_ssg_2026,
title = "Structured Spectral Graph Representation Learning for Multi-label Abnormality Analysis from 3D CT Scans",
author = "Di Piazza, Theo and Lazarus, Carole and Nempont, Olivier and Boussel, Loic",
journal = "Machine Learning for Biomedical Imaging",
volume = "2026",
issue = "June 2026 issue",
year = "2026",
pages = "359--388",
issn = "2766-905X",
doi = "https://doi.org/10.59275/j.melba.2026-87e3",
url = "https://melba-journal.org/"
}
```