EAGLE

EAGLE (Efficient Approach for Guided Local Examination) is a computational pathology framework from the Kather Lab at TU Dresden that turns a gigapixel whole-slide image into a single slide-level embedding by examining only a handful of regions. Conventional pipelines encode every tile in a slide — roughly 18,000 tiles per slide at 0.5 µm/pixel in the cohorts studied here — and then train a separate attention-based aggregator for each downstream task. EAGLE instead emulates how a pathologist works: survey the slide, identify the few fields that carry the diagnostic signal, and look at those closely.

The framework is deliberately built from components that already exist rather than a new pretrained encoder. It chains CHIEF, used frozen and task-agnostically to rank tiles by attention, with Virchow2, used to extract high-resolution features from the tiles CHIEF selects. Because both backbones stay fixed and the pooling step is a plain average, the resulting embedding is deterministic and reusable: extract it once per slide, then fit whatever lightweight classifier a given task needs.

EAGLE was released as a preprint in February 2025 and published in Nature Communications in 2026, alongside a GPL-3.0 extraction script that runs the pipeline on new slides.

#Key Features

• Twenty-five tiles per slide: Tile-count ablations showed accuracy rising up to 25 tiles and flattening beyond it, so 25 became the default budget — small enough that every region contributing to a prediction can be displayed and audited.
• Frozen, task-agnostic selection: CHIEF attention scores are computed once per slide and never refit per task, which is what makes a single embedding valid across an entire benchmark suite.
• Unweighted mean pooling: Averaging the 25 Virchow2 embeddings performed on par with attention-weighted pooling, and equal weighting was adopted because it keeps region-level contributions transparent.
• Auditable by construction: Systematic negative controls, random-selection baselines, and attention-concentration (Gini) analyses accompany the benchmark, and the exact tiles behind each prediction can be overlaid on the slide.

#Technical Details

The pipeline runs in three steps, all at 2 µm/pixel: CTransPath embeddings are extracted for every tile, CHIEF converts them into per-tile attention scores, and the top 25 tiles are re-extracted with Virchow2 and averaged into one 1280-dimensional patient embedding. On an L40 GPU the three steps take 2.01 s, 0.36 ms, and 0.26 s respectively — 2.27 s per slide, over 99% faster than encoding every tile at the conventional 0.5 µm/pixel. Only about 2% of a slide's tiles are re-encoded in detail. The core benchmark comprises 31 tasks across breast, colorectal, gastric, and non-small cell lung cancer, with five-fold cross-validated classifiers trained on TCGA and evaluated on fully external cohorts (CPTAC, DACHS, Kiel, Bern, IEO). EAGLE reached the highest mean AUROC at 0.742, ahead of TITAN at 0.740 and Virchow2 aggregated through STAMP at 0.723, and led on AUPRC (0.566) with COBRA second. It exceeded 0.800 AUROC on 39% of tasks versus 35% for TITAN. The full evaluation spans 43 tasks across nine cancer types, adding treatment-response endpoints, and reports gains of up to 23% over patch-aggregation baselines including CONCH, Prov-GigaPath, and MADELEINE.

#Applications

EAGLE targets biomarker prediction and prognostic modeling in clinical pathology workflows — inferring molecular alterations, protein expression, or treatment response directly from H&E slides. The efficiency changes what is practical: a slide can be embedded on modest hardware in seconds rather than minutes, making real-time and point-of-care deployment feasible and removing the dependence on high-performance computing that keeps many groups out of computational pathology. Because every patient reduces to one fixed vector, the same embeddings also support fast slide retrieval across archives and joint analysis with genomic or transcriptomic data.

#Impact

EAGLE's contribution is a demonstration that selective examination beats exhaustive encoding: a frozen two-model chain over 25 tiles matches or exceeds purpose-trained slide encoders while discarding 99% of the compute. Its practical value lies as much in auditability as accuracy, since a prediction traceable to 25 displayable tiles is far easier to review than one aggregated over thousands. The limits are stated plainly by the authors: the mean AUROC of 0.742 is not yet adequate to replace expert assessment, downstream classifiers are still fit per task, and the framework inherits whatever biases CHIEF and Virchow2 carry from their own pretraining corpora.