Seeing Beyond the Lesion: Disease Recognition from Reactive CNS Tissue

TL;DR - Weakly supervised pathology models can predict underlying CNS diseases from reactive, non-lesional biopsy tissue after controlling for sampling-related shortcuts. The findings suggest conventionally non-diagnostic tissue contains subtle disease signals and highlight provenance-only baselines as essential controls.

• Four pathology foundation models were evaluated as frozen patch encoders on 245 whole-slide images from 186 patients.
• Coarse disease categories were largely predictable from slide size, exposing a major acquisition confound.
• For three finer diagnostic distinctions, predictions remained above chance after removing this confound (permutation testing, p ≤ 10⁻⁴).
• Performance did not significantly differ among the four encoders; contribution maps and expert review assessed whether signals arose from reactive tissue rather than artifacts such as blood.