CT-RATE Image Classification Result

[Unstoppableljq]

Thank you for your work.

I would appreciate it if you could provide some details on the evaluation method for image classification in CT-Rate. Is it based on linear probing?

[srikantak-221]

Hey @Unstoppableljq ,
No, MedGemma does not use linear probing for image classification evaluation. All evaluations for MedGemma were performed as “single inference runs per example”, with temperature adjustments applied for medical benchmarks.
While MedGemma does use the MedSig encoder (the same backbone used in MedSigLIP), the evaluation strategy differs, unlike MedSigLIP, which uses linear probing to assess frozen feature embeddings, MedGemma measures end-to-end model performance directly.
Please refer here for more details on Evaluations.
Thank you!