Pyramid-level training with stain normalization for breast cancer lymph-node metastasis segmentation in whole-slide images

Whole-slide histopathology images encode metastatic tissue information simultaneously at multiple spatial scales, from fine cellular morphology at high magnification to broad tissue architecture at low magnification. Existing segmentation models commonly operate at a single fixed resolution, leaving complementary information distributed across the whole-slide image pyramid underused. We present a multilevel segmentation framework with stain-normalization preprocessing that trains a unified model jointly across four study-defined WSI pyramid views (5×-40×), integrating Vahadane stain normalization and structured class-balanced sampling stratified by pyramid level, dataset, and tissue class. Applied to a 280-slide CAMELYON16/CAMELYON17 study cohort under strict slide-level data separation (196 training, 42 validation, 42 held-out test slides), the framework achieves a Dice coefficient of 92.39% on the held-out test set. Relative to the single-scale U-Net baseline (89.00%), the paired slide-level improvement is 3.39 percentage points (95% bootstrap CI: 3.19-3.60 pp; Cohen's dz=4.98; two-sided permutation p < 0.00001). Per-level evaluation confirms that full pyramid fusion outperforms each individual scale, while a boundary-focused L1 audit yields a mean HD95 of 53.88 pixels (95% CI: 50.28-57.64). The contribution is therefore a training-sampling and evaluation pipeline rather than a new network layer or architecture. These results show that multiresolution training, stain normalization, structured sampling, and leakage-controlled evaluation can materially improve breast lymph-node metastasis segmentation on established WSI benchmarks. A supplementary external cross-dataset evaluation on 103 paired BCSS breast-histopathology ROIs, performed without retraining or fine-tuning, yielded a pooled Dice of 70.58%, IoU of 54.54%, precision of 55.95%, and recall of 95.57%; the mean per-ROI Dice was 67.89% (95% bootstrap CI: 65.00-70.76%). These external results demonstrate substantial sensitivity to tumor-like tissue under a different image source and semantic annotation scheme, while also exposing the expected effect of annotation granularity and class-definition mismatch on overlap and boundary metrics.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-13
DOI
https://doi.org/10.1371/journal.pone.0362258
Primary Topic
AI in cancer detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pyramid-level training with stain normalization for breast cancer lymph-node metastasis segmentation in whole-slide images

Ümit Şentürk, A. Yılmaz, Selma Erdoğan Düzcü, Alperen Akgül et al.
1 links
PLoS ONE
AI in cancer detection
article

Pyramid-level training with stain normalization for breast cancer lymph-node metastasis segmentation in whole-slide images

Ümit Şentürk, A. Yılmaz, Selma Erdoğan Düzcü, Alperen Akgül, Hamza Almobayed
article en

Abstract

Whole-slide histopathology images encode metastatic tissue information simultaneously at multiple spatial scales, from fine cellular morphology at high magnification to broad tissue architecture at low magnification. Existing segmentation models commonly operate at a single fixed resolution, leaving complementary information distributed across the whole-slide image pyramid underused. We present a multilevel segmentation framework with stain-normalization preprocessing that trains a unified model jointly across four study-defined WSI pyramid views (5×-40×), integrating Vahadane stain normalization and structured class-balanced sampling stratified by pyramid level, dataset, and tissue class. Applied to a 280-slide CAMELYON16/CAMELYON17 study cohort under strict slide-level data separation (196 training, 42 validation, 42 held-out test slides), the framework achieves a Dice coefficient of 92.39% on the held-out test set. Relative to the single-scale U-Net baseline (89.00%), the paired slide-level improvement is 3.39 percentage points (95% bootstrap CI: 3.19-3.60 pp; Cohen's dz=4.98; two-sided permutation p < 0.00001). Per-level evaluation confirms that full pyramid fusion outperforms each individual scale, while a boundary-focused L1 audit yields a mean HD95 of 53.88 pixels (95% CI: 50.28-57.64). The contribution is therefore a training-sampling and evaluation pipeline rather than a new network layer or architecture. These results show that multiresolution training, stain normalization, structured sampling, and leakage-controlled evaluation can materially improve breast lymph-node metastasis segmentation on established WSI benchmarks. A supplementary external cross-dataset evaluation on 103 paired BCSS breast-histopathology ROIs, performed without retraining or fine-tuning, yielded a pooled Dice of 70.58%, IoU of 54.54%, precision of 55.95%, and recall of 95.57%; the mean per-ROI Dice was 67.89% (95% bootstrap CI: 65.00-70.76%). These external results demonstrate substantial sensitivity to tumor-like tissue under a different image source and semantic annotation scheme, while also exposing the expected effect of annotation granularity and class-definition mismatch on overlap and boundary metrics.

PLoS ONEVol. 21(11)
Bolu Abant İzzet Baysal University (TR)
Good health and well-being
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon