Rethinking Slide-Level Adaptation of Pathology Foundation Models

The emergence of pathology foundation models (PFMs) has transformed histopathological image analysis, replacing conventional image encoders pretrained on natural images and becoming the mainstream feature extractor for patches in whole slide images (WSIs). To perform slide-level analysis, multiple instance learning (MIL) remains the dominant paradigm. However, such approaches rely on task-specific slide-level representations, which often lack generality and carry a risk of feature overfitting. Another simple way is to directly aggregate patch features into a slide representation via training-free pooling methods. Nevertheless, before the advent of PFMs, such approaches generally performed poorly. In this work, we demonstrate that task-agnostic slide representations, aggregated via mean pooling of PFM-derived patch features, can successfully support a wide range of WSI classification simply by using a deeper multilayer perceptron. For instance, this method outperforms popular MIL-based methods on large-scale pan-cancer classification and demonstrates competitive few-shot learning ability, robustness, and transferability. This finding encourages rethinking the slidelevel adaptability of PFMs and shifting future research from designing complex aggregators toward developing stronger pretraining and more adaptable feature learning.