ELF

Benchmark studies of pathology foundation models keep returning the same awkward result: no single encoder wins everywhere, and the ranking reshuffles not only between tasks but between datasets for the same task. A group choosing one tile encoder for a new biomarker study is making a bet it has no way to settle in advance, and the obvious remedy — distilling several encoders into one averaged model — tends to blur the very differences that made them worth combining.

ELF (Ensemble Learning of Foundation models) declines to choose. Built in Ruijiang Li's laboratory at Stanford University with collaborators at Memorial Sloan Kettering Cancer Center, it passes a whole-slide image through five frozen tile-level encoders — Prov-GigaPath, CONCHv1.5, Virchow2, H-optimus-0 and UNI — and learns a single shared slide encoder that aggregates each encoder's tile embeddings separately, then concatenates the five slide vectors into one representation. The acronym is unrelated to the ELF/ETS transcription factor family.

Because the expensive pretraining sits in the tile encoders, ELF's own contribution is a compact aggregator applied from a fixed checkpoint and kept frozen downstream. That is what makes it usable on the cohorts precision oncology actually has: a few dozen patients on one drug regimen, far too few to fine-tune a tile-level model against. It is benchmarked against the slide-level models CHIEF, TITAN and Prov-GigaPath, and against COBRA as an aggregation baseline.

#Key Features

• One encoder, five feature streams: the same slide encoder reads embeddings from all five tile models, so the ensemble is learned rather than voted, and the concatenated output preserves each model's native feature space instead of averaging them away.
• Complementarity measured, not assumed: pairwise correlations between the attention maps the five streams produce are consistently low, evidence that the constituent models weight different regions of the same slide.
• Data-efficient downstream use: with the slide encoder frozen, response models were fit on cohorts as small as 36 patients, where tile-level fine-tuning is not an option.
• Cheap aggregation: at 14.3 million parameters the slide encoder is smaller than COBRA II (15.5M), TITAN (48.5M) and Prov-GigaPath (86.3M), and adds roughly 0.1 s per slide.
• Attention heatmaps over the slide: the aggregator's weights project back onto the image, letting a pathologist inspect the regions driving a prediction.

#Technical Details

The slide encoder is a gated attention-based multiple instance learning (ABMIL) model with eight heads. Tile embeddings of 768, 1024, 1280 or 1536 dimensions are linearly interpolated to a common 768 dimensions, LayerNorm-normalized, then reshaped so each head attends over its own 96-dimensional slice; head attention scores are averaged, softmax-normalized and used to pool both the unified features and the encoder's native-dimension features. Pretraining used MoCo v3-style momentum contrastive learning (temperature 0.2) to align slide embeddings across models and across views, plus weakly supervised cancer-detection and organ-classification heads, for 300 epochs on eight NVIDIA H100 GPUs at an effective batch size of 768. The corpus was 53,699 whole-slide images across 20 anatomical sites from eleven public collections including TCGA, GTEx, PANDA, PAIP and ACROBAT, tiled at 10x magnification with CLAM.

On disease subtyping ELF reached balanced accuracy of 0.757 on the 30-class EBRAINS brain tumor task and 0.457 on 7-class BRACS breast subtyping, against 0.393 for TITAN. Across 84 biomarker-indication combinations in TCGA it averaged AUC 0.719 ± 0.142, above TITAN (0.701), CHIEF (0.676) and Prov-GigaPath (0.678), and reached mean AUC 0.886 ± 0.015 for microsatellite instability status across four colorectal cohorts.

#Applications

The intended use is feature extraction for retrospective computational pathology studies where labeled slides are scarce. Across four chemotherapy and targeted-therapy cohorts spanning breast and ovarian cancer, ELF scores averaged AUC 0.759 for response, and the derived risk scores separated progression-free and recurrence-free survival. On a real-world immunotherapy dataset of 13 cohorts and 1,057 patients across eight cancer types, it averaged AUC 0.724 for durable checkpoint-inhibitor response; in an external MSKCC non-small-cell lung cancer cohort, combining the ELF score with PD-L1 and tumor mutational burden reached AUC 0.722, ahead of either clinical marker alone.

#Impact

ELF argues that the proliferation of pathology foundation models is a resource rather than a problem to be settled by picking a winner, and that the useful integration point is a slide-level encoder trained over their outputs. The cost is operational: inference requires running all five constituent encoders, about 14 minutes per slide sequentially or 3.5 minutes with the extractors parallelized across GPUs, and each constituent carries its own license and access terms. The Stanford and MSKCC immunotherapy cohorts are not public, so those results cannot be reproduced from the released artifacts. The authors frame the work as supporting further evaluation of ensemble learning for pathology rather than as a validated clinical tool, and the released code and weights are offered for non-commercial academic use.