VFSS_analysis

This repository provides tools to preprocess Videofluoroscopic Swallowing Studies (VFSS), automatically label them using a deep learning–based framework, and extract dynamic parameters characteristic of dysphagia.

Please download the models from: https://doi.org/10.5281/zenodo.17191973
and save them in the models folder.

Citation

If you use this repository in your research, please cite our publication:

Cubero L, Tessier C, Castelli J, Robert K, de Crevoisier R, Jégoux F, Pascau J, Acosta O.
Automated dysphagia characterization in head and neck cancer patients using videofluoroscopic swallowing studies.
Comput Biol Med. 2025 Mar;187:109759. doi: 10.1016/j.compbiomed.2025.109759. Epub 2025 Feb 6. PMID: 39914196.
(https://doi.org/10.1016/j.compbiomed.2025.109759)

For any issues or questions, please feel free to contact us: lcubero@ing.uc3m.es ; jpascau@ing.uc3m.es ; acosta@univ-rennes.fr

---

Installation

Follow these steps to set up the repository:

git clone https://github.com/BSEL-UC3M/VFSS_analysis.git
cd VFSS
conda env create -f environment.yml
conda activate VFSS_env
pip install -e .

Notes & Considerations

• This repository is under active development.
• The current version of nnU-Net used for training and inference is version 1. An update to version 2 is planned.
• Two models are provided (and can be downloaded from the Zenodo repository indicated above). These are the two best-performing models: Task010_VFSS (default), and Task008_VFSS
  • Be aware that the performance on the models will depend on your data (and how similar it is to the training data). All the preprocessing is adapted to the training data from our institution. Modifications may be required.
• A sample VFSS file is provided at:
  VFSS_analysis/data/raw_VFSS/test/healthy_001
  • The file is in AVI format, compressed to preserve quality. It may not be compatible with all video players.
  • The VFSS with the manual labels, as well as the predictions and parameters analyzed, are also provided.

---

Usage

This repository provides a full pipeline for analyzing new VFSS (Videofluoroscopic Swallowing Studies).
The main script processes raw VFSS videos by:

1. Preprocessing

   • Reads the input VFSS video (.avi format).
   • Converts frames into images.
   • Resizes and formats them for nnU-Net inference.

2. Inference

   • Runs nnU-Net (v1) segmentation models. ··· You can choose between the two best-performing models: Task010_VFSS (default), and Task008_VFSS
   • Produces predicted region-of-interest (ROI) masks.
   • Applies postprocessing to clean predictions.

3. Labeled Video Generation

   • Creates videos with the predicted labels overlaid on the original VFSS.
   • Saves them per patient and time point.

4. Parameter Extraction

   • Computes dynamic parameters (areas and landmarks) from the segmented ROIs.
   • Saves them in structured output files for downstream analysis.