The official implementation of Few-Shot, Now for Real: Medical VLMs Adaptation without Balanced Sets or Validation.
📜 Medical Image Computing and Computer Assisted Intervention (MICCAI)
Julio Silva-Rodríguez¹, Fereshteh Shakeri^1,2, Houda Bahig², Jose Dolz^1,2, Ismail Ben Ayed^1,2
1ÉTS Montréal ⋅ ² Centre Hospitalier de l’Université de Montréal (CRCHUM)
| Conference | ArXiv |

• Install in your environment a compatible torch version with your GPU. For example:

conda create -n sstext python=3.11 -y
conda activate sstext
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124

git clone https://github.com/jusiro/SS-Text.git
cd SS-Text
pip install -r requirements.txt

• Configure data paths (see ./local_data/constants.py).
• Download, and configure datasets (see ./local_data/datasets/README.md).

We present the basic usage here.

(a) Features extraction:

• python extract_features.py --task MESSIDOR

(b) Standard adaptation:

• python adapt.py --task MESSIDOR --k 4 --scenario standard --adapt SStext+

(c) Relaxed adaptation scenario:

• python adapt.py --task MESSIDOR --k 4 --scenario relaxed --adapt SStext+

(d) Realistic adaptation scenario:

• python adapt.py --task MESSIDOR --k 4 --scenario realistic --adapt SStext+

You will find the results upon training at ./local_data/results/.

If you find this repository useful, please consider citing the following sources.

• For the original SS-Text publication (LINK), designed for fast conformal prediction:

@inproceedings{fca25,
    title={Full Conformal Adaptation of Medical Vision-Language Models},
    author={Julio Silva-Rodríguez and Leo Fillioux and Paul-Henry Cournède and Maria Vakalopoulou and
    Stergios Christodoulidis and Ismail {Ben Ayed} and Jose Dolz},
    booktitle={Information Processing in Medical Imaging (IPMI)},
    year={2025}
}

• For SS-Text+, designed for imbalanced scenarios, and realistic few-shot settings:

@inproceedings{sstext25,
    title={Few-Shot, Now for Real: Medical VLMs Adaptation without Balanced Sets or Validation},
    author={Julio Silva-Rodríguez and Fereshteh Shakeri and Houda Bahig and Jose Dolz and Ismail {Ben Ayed}},
    booktitle={Medical Image Computing and Computer Assisted Intervention (MICCAI)},
    year={2025}
}