ryhm/siglip_single_cell_blindspots

SigLIP Single-Cell Blind Spots

This dataset contains 10 images from the Acevedo single-cell dataset where the SigLIP base model fails in single-cell classification.

Dataset Structure

Each entry contains:

• image: filename of the image
• ground_truth: actual class of the cell
• model_output: class predicted by SigLIP

Reproducibility / Experiments

All experiments related to this dataset were conducted in the siglip_2.ipynb notebook.

• The notebook includes:
  • Loading the SigLIP model (google/siglip-base-patch16-224)
  • Zero-shot single-cell classification
  • Analysis of blind spots
  • Visualizations of predictions
• You can open and run the notebook in Google Colab to reproduce the results.

Observations

SigLIP struggles with single-cell classification. The dataset highlights these blind spots.

Suggested Fine-Tuning

To improve model performance:

• Use DINOBloom as a foundation model.
• Fine-tune on multiple single-cell datasets including Acevedo.

References

• Model tested: google/siglip-base-patch16-224