ECTIL

The density of stromal tumour-infiltrating lymphocytes on an H&E slide is an established prognostic factor in triple-negative breast cancer, and it is scored by eye. Automating that score has been attempted many times, but the published computational TIL assessment pipelines are elaborate: they chain a tissue-segmentation network, a cell-detection network, and a rule-based aggregation step, and training them requires tens of thousands of pixel-level and cell-level annotations from expert pathologists. That annotation cost is what has kept such models out of routine use.

ECTIL, developed at the Netherlands Cancer Institute by Yoni Schirris and colleagues with the University of Amsterdam and University Medical Center Utrecht, removes the intermediate steps entirely. Tile features are extracted once by a frozen self-supervised pathology encoder, and a small attention-based head regresses the slide's TIL percentage straight from them — the same frozen-backbone-plus-thin-head pattern that UNI and CONCH were built to enable, applied here to a single continuous clinical readout. Training needs only a few hundred slide-level scores, two orders of magnitude fewer labels than the segmentation-based alternatives, and finishes in about ten minutes on a CPU.

The claim the paper actually tests is generalisation. A model trained on 356 TCGA slides is applied unchanged to five independent external cohorts spanning three countries and including three randomised clinical trials — 2,340 patients in total, 790 of them with triple-negative disease. It was published in The Lancet Digital Health in November 2025.

#Key Features

• Single-step regression: No segmentation mask, no cell detector, no post-hoc rules — the TIL percentage is a direct output, which removes the failure modes that compound across a multi-stage pipeline.
• Two orders of magnitude fewer labels: Supervision is one number per slide. The comparators require roughly 85,000 cell annotations and 300 densely annotated regions.
• Cross-cohort validation: Five external cohorts, three of them randomised trials, scanned on different scanners in different countries, with no per-cohort retuning.
• Independent prognostic value: In multivariate Cox regression on the PARADIGM cohort, each 10% increment of the ECTIL-combined score carried a hazard ratio of 0.85 (95% CI 0.77-0.93; p=0.0007) for overall survival, against 0.86 (0.81-0.92) for the pathologist.
• Inspectable predictions: Per-tile score and attention heatmaps let a pathologist see which regions drove a given slide's score.

#Technical Details

Features come from RetCCL, a self-supervised convolutional pathology encoder, which produces a 2048-dimensional vector per 512x512-pixel tile at 0.5 microns per pixel. The head is a gated attention-based multiple instance learning model: a ReLU layer projecting 2048 to 512 dimensions, a gated attention module with a 128-dimensional hidden layer, and a linear layer with a sigmoid mapping the pooled representation to [0,1]. Three variants were trained — ECTIL-TCGA on 356 TCGA slides with folds split by medical centre, ECTIL-TNBC on 400 triple-negative slides from four cohorts, and ECTIL-combined on 1,964 slides from five cohorts with leave-one-cohort-out folds.

ECTIL-TCGA reached Pearson correlations of 0.54-0.74 and AUROCs of 0.80-0.94 against the pathologist across the five external cohorts, at a clinically used TILs-high threshold of 30%, peaking on MATADOR (r=0.74, AUROC 0.94). On the held-out PARADIGM cohort, ECTIL-combined reached r=0.69 and AUROC 0.85, ahead of ECTIL-TNBC (0.64, 0.83) and ECTIL-TCGA (0.58, 0.80). Discordant-case review by a board-certified pathologist identified where the model diverges: it performs worse on invasive lobular carcinoma (r=0.36, AUROC 0.64), where tumour cells are scored as lymphocytes, and it attends to lymphovascular invasion, fibrotic regions, and lymphocytes at the tumour boundary that a pathologist would exclude. Its predicted range is also compressed relative to the pathologist's.

#Applications

The intended use is pre-screening: identifying patients whose TIL levels make them candidates for prospective chemotherapy de-escalation trials in low-stage triple-negative breast cancer, and assisting pathologists during diagnostic work-up, where manual scoring is slow and inter-observer variability is well documented. Because the trained head is small and the inference script accepts arbitrary whole-slide images, translational groups can run it over an existing retrospective cohort cheaply, and retrain it on their own scores if local scanners or scoring conventions differ.

#Impact

ECTIL is a demonstration that in computational pathology, moving complexity out of the task-specific model and into a frozen general-purpose backbone can improve generalisation rather than cost it. The design also makes clinical iteration practical: a ten-minute retraining cycle on slide-level labels is something a hospital group can actually run. Code is released under Apache-2.0 with model weights in a public model zoo, and the pathologist TIL scores collected for the TCGA breast cohort are published alongside as a reusable dataset. Validation remains retrospective, and the prognostic analysis rests on a single cohort of systemic-treatment-naive patients under 40 with node-negative triple-negative disease.