Reproduce “MI-VisionShot: Few-shot adaptation of vision-language models for slide-level classification of histopathological images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pablo Meseguer, Rocío del Amor, Valery Naranjo
The MI-VisionShot framework introduces a training-free method for adapting vision-language models to perform slide-level classification of histopathological images in few-shot learning scenarios. This approach reduces the variability of previous zero-shot methods, achieving up to 79.9% balanced accuracy and outperforming baselines by up to 8.2% in low-shot settings.
Key insights
- Text-guided selection of the top-K most discriminative patches within WSIs is crucial for robust class prototype construction in few-shot histopathological classification.
- Integrating linguistic knowledge from VLMs through a training-free, prototype-based MIL approach significantly mitigates the prompt dependence and variability of prior zero-shot methods.
- Focusing on relevant visual information from gigapixel images allows for improved class separation in the latent space, leading to more effective classification with minimal data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.