TCLA: Training-Free Class-wise Logit Adaptation for Medical Vision-Language Models

Researchers present TCLA, a purely training-free few-shot adaptation method designed to improve the performance of Medical Vision-Language Models (VLMs) on out-of-distribution data. The method corrects inference logits using a small set of support samples to boost inter-class deconfusion and reduce domain shift without introducing trainable components.

• TCLA is model-agnostic and fast, addressing instability issues found in existing few-shot methods that rely on additional trainable parameters.
• It improves performance across nine datasets spanning multiple medical imaging modalities, including X-ray, Ultrasound, MRI, CT, and Histopathology.
• In most cases, TCLA outperforms existing training-based adaptation methods while maintaining robustness on different medical data types.

This approach offers a stable alternative for adapting pretrained VLMs in extremely low-data regimes where traditional few-shot methods may fail.