gi-siglip2-dino-hyperkvasir-checkpoints

GI-SigLIP2-DINO-HyperKvasir Checkpoints

Selected checkpoints and evaluation results from the GI-SigLIP2-DINO-HyperKvasir++ project.

Dataset

Used dataset:

• Sahibnoor1/gi-endoscopy-megabank-stage2

Included checkpoints

finetune_from_ssl_step_12500/best.pt

Best SSL fine-tuned checkpoint by validation Macro F1.

Validation:

• Macro F1: 0.679335
• Accuracy: 0.903327
• MCC: 0.895075

Test:

• Macro F1: 0.596419
• Micro F1: 0.892028
• MCC: 0.882982

finetune_from_ssl_40k/best.pt

Fine-tuned model from final 40k SSL checkpoint.

Validation:

• Macro F1: 0.674056
• Accuracy: 0.898933
• MCC: 0.890506

specialist_polyp_dye/best.pt

Specialist classifier for:

• polyps
• dyed-lifted-polyps
• dyed-resection-margins

Validation:

• Accuracy: 0.969027
• Macro F1: 0.968667
• MCC: 0.953643

Notes

Full DINO/iBOT-style SSL was run on the full 1.125M-image GI-Endoscopy MegaBank Stage 2 dataset.

The best final model from earlier supervised training remains the supervised SigLIP2-224 checkpoint unless a routed specialist system improves validation and test metrics.