pelvis-siglip-medsiglip-sequence

Pelvis-SigLIP — MedSigLIP (Sequence)

Fine-tuned MedSigLIP vision backbone + linear head for pelvic MRI sequence classification, from the MICCAI CAPI WOMAN 2026 workshop paper "Pelvis-SigLIP: Benchmarking Vision-Language Models for Female Pelvic MRI Series Retrieval across Zero-Shot, Linear-Probe, and Fine-Tuning."

• Classes (8): ADC, DCE, DWI, T1, T1FS, T2, T2FS, other
• Input: a single 2D slice (central slice of the series), 448×448, normalised with mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]
• Training data: 2,702 pelvic MRI series from 534 patients across 5 public cohorts, multi-vendor (Siemens/GE/Philips), 1.5T & 3.0T
• Selection: best of 5 patient-level cross-validation folds by held-out test accuracy (fold 3 kept; all 5 shown below)

Test accuracy

Mean across folds: val = 0.9641, test = 0.9512. (The paper's Table 2 reports the mean ± std across all 5 folds' own models; this released checkpoint is the single best-performing fold, not an ensemble.)

Usage

See Pelvis-SigLIP on GitHub for the loading code and a runnable example starting from a raw .nii / .nii.gz scan: https://github.com/MaximilianLindholz/Pelvis-SigLIP

Intended use

Metadata-independent MRI series retrieval / classification — identifying sequence from the image alone when DICOM series descriptions are missing, inconsistent across vendors, or corrupted (e.g. after a PACS migration). Trained and evaluated on female pelvic MRI only — not validated on other anatomy.

Limitations

• Evaluated as classification accuracy, not ranked retrieval (precision@k) over a full PACS-scale index.
• Trained on 5 public research cohorts; performance on your local scanner/protocol distribution may differ.
• No per-class error analysis reported (see paper limitations).

Citation

@inproceedings{lindholz2026pelvissiglip,
  title     = {Pelvis-SigLIP: Benchmarking Vision-Language Models for Female Pelvic MRI Series Retrieval across Zero-Shot, Linear-Probe, and Fine-Tuning},
  author    = {Lindholz, Maximilian and Ruppel, Richard and Hamm, Charlie Alexander and Kn{\\"u}pfer, Anika and Eminovic, Semil and Schmidt, Robin and Haa
ck, Anna-Maria and El-Nahry, Yasmin and Aleixo, Carolina and Hutter, Jana and Mechsner, Sylvia and Penzkofer, Tobias},
  booktitle = {Proceedings of the MICCAI 2026 Workshop CAPI-WOMEN},
  year      = {2026},
  note      = {Oral}
}