TCLA: Training-Free Class-wise Logit Adaptation for Medical Vision-Language Models
TCLA (Training-Free Class-wise Logit Adaptation) is a training-free class-wise logit adaptation method designed to improve the zero-shot performance of medical vision-language models (VLMs) on out-of-distribution (OOD) data. It introduces no additional trainable components and aims to address domain shift and class bias.
TCLA is a training-free class-wise logit adaptation method that enhances the zero-shot performance of medical vision-language models on out-of-distribution data without introducing additional trainable components.
TCLA mitigates domain shift and class bias through class-wise logit adaptation without requiring additional training, potentially improving OOD generalization by adjusting the logit distribution.
The performance degradation of medical VLMs on OOD data is a pain point in the industry. TCLA offers a lightweight solution that may facilitate the practical deployment of medical image analysis.
TCLA improves model robustness without training, reducing the deployment cost of medical AI systems and helping accelerate clinical adoption.
If TCLA proves effective on more medical datasets, it could become a standard post-processing step for medical VLMs.