Skip to main content
Back to timeline
arXivSource publication:

Tissue detection stops at average overlap: SCOUT cuts case groups needing mask correction from 38.1% to 18.6%

Synopsis

Tissue detection in whole-slide images is usually judged by average overlap, but that average hides missed pale tissue, small fragments and staining variation. SCOUT first builds an initial mask from a thumbnail, then a lightweight audit network estimates where errors concentrate and sends only those high-risk windows to a frozen pathology foundation model, merging the results with an asymmetric consensus rule. On a public multi-source, multi-stain benchmark, the share of case groups needing mask correction falls from 38.1% for the best released detector to 18.6%, with the clearest gains on immunohistochemistry.

Source-provided article image: SCOUT: Risk-Guided Foundation-Model Escalation for Robust Tissue Detection in Whole-Slide Images
arXiv

Interpretation

Tissue-detection reliability is now counted as whether a mask needs correction, not only as average overlap with manual annotation. Most pipelines have relied on thresholding or pretrained detectors, judged by mean overlap on H&E; on a multi-stain public benchmark, released detectors still leave many case groups needing correction even after benchmark-specific post-processing tuning. On the Tissues with Issues test set, SCOUT lowers the share of case groups needing correction to 18.6%, against 38.1% for the development-tuned AtlasPatch, a paired difference of -19.5 percentage points with a 95% confidence interval of [-29.2, -9.7].

The foundation model is invoked only in high-risk windows, while the rest keeps the thumbnail detector's output. The audit network is a U-Net with a MobileNetV3-small encoder and about 1.4M parameters; it does not output a replacement mask but estimates where the initial mask misses tissue or marks background as tissue. The thumbnail is split into 448 µm windows and only the highest-risk 50% of candidates go to the frozen Virchow2 encoder, whose representations a lightweight decoder maps back to tissue probabilities; ablations show random window selection gives limited gain, while processing every candidate yields no further reduction in failures.

The merge rule sets different thresholds for adding tissue and for keeping tissue, to suppress false positives from the foundation model. The foundation model can misclassify unstained glass and debris as tissue, so it cannot simply replace the initial mask. Inside a selected window, keeping an existing positive pixel requires a lower foundation-model probability, while turning a background pixel into tissue requires the higher probability of 0.5; components smaller than 0.01 mm² are then removed and unscanned regions excluded.

Perspective

Evaluation rests on one public multi-source, multi-stain benchmark plus a small external slide set; case groups do not always correspond to patients, and ablations also use the test set. Downstream validation uses the 500 CAMELYON17 lymph-node training slides with leave-one-centre-out folds.

Reference masks lack expert adjudication, and the failure thresholds are not validated against clinical consequences; much of the source-held-out gain arises on ACROBAT, and overlap with foundation-model pretraining cannot be excluded. The review guarantee is marginal over the benchmark population and does not control per-site or accepted-slide failure rates; SCOUT requires a GPU and does not achieve the lowest HD95.