Cross‑site lung histopathology classification must handle stain variation, non‑IID client data, missing classes, and the cost of adapting large pathology encoders. This paper evaluates FedHisto‑PAST v2 for three‑way classification of adenocarcinoma (ACA), Normal tissue, and squamous cell carcinoma (SCC).
FedHisto‑PAST v2 builds on a frozen HIBOU‑B foundation model and introduces parameter‑efficient adaptation layers. Key techniques include stain‑conditioned paired‑view prediction with feature consistency, reliability‑aware prototype learning, and adaptive federated aggregation.
Experiments were conducted in a five‑client, non‑IID, raw‑data‑local simulation with a fixed internal evaluation set, client‑level analysis, component ablations, communication accounting, and an exploratory LungHist700 cohort influenced by development decisions. All principal methods achieved near‑ceiling performance on the internal split, limiting discrimination there.
On LungHist700, FedHisto‑PAST v2 attained a Macro‑F1 of 0.728560 and a balanced accuracy of 0.730454. Recognition of Normal and SCC improved while ACA recall decreased, and calibration remained imperfect. Ablation on external data showed that only prediction‑level consistency contributed a clear independent gain; feature consistency and prototype regularization did not yield conclusive overall Macro‑F1 improvements. The framework updated merely 1.253841% of model parameters.
The results provide exploratory cross‑dataset evidence for stain‑aware, parameter‑efficient federation, but do not establish formal privacy guarantees, patient‑level independence, prospective deployment, or clinical validation.
Review: The study demonstrates that minimal‑parameter adaptation can enable effective cross‑site pathology classification under resource constraints, offering a promising direction for future privacy‑preserving and clinically viable federated solutions.