Normal Breast Tissue (NBT)-Classifiers: Advancing Compartment Classification in Normal Breast Histology

Key results

1. ShownBoth final classifiers (512 px and 1024 px inputs) generalise to three held-out cohorts, one of them scanned at 20x. Accuracy shows no bias by patient age or by tissue source.

   Class AUCs pooled over KHP, EPFL and SGK: 512px model 0.98 epithelium, 0.98 stroma, 1.00 adipocytes; 1024px model 0.99, 0.98, 1.00. Cohort-specific AUCs varied somewhat. Confounder checks found no bias: age used 15 external patients in three age bands; source used 30 NKI/BCI (training-cohort) patients across three sources; each check was repeated three times. · NBT-Classifiers demonstrate strong…; External validation; Confounder analysis

2. ShownOn the same external patches, the 1024px NBT-Classifier matched or beat HistoROI, a ResNet18 model trained on mixed breast tissue from normal to cancerous. The gap was largest for adipocytes.

   AUC, NBT-Classifier vs HistoROI: epithelium 0.99 vs 0.99, stroma 0.98 vs 0.97, adipocytes 1.00 vs 0.95 (DeLong's test, P<0.0001). Class accuracy: 87%, 94%, 96% vs 80%, 91%, 72% for epithelium, stroma and adipocytes. · Training exclusively on normal tissue enables…; Implementation of HistoROI

3. SuggestedAlthough it never saw abnormal tissue, the NBT-Classifier's epithelium probabilities for normal patches are shifted relative to benign, precancerous and cancerous patches. It separates normal from abnormal epithelium better than HistoROI, which the authors read as learning normal-specific features.

   Test set: BRACS, 56,000 patches of 1024 × 1024 px at 40x, 8,000 per class (N, PB, UDH, FEA, ADH, DCIS, IC; Methods lists only N, ADH, DCIS, IC). The NBT-Classifier called 96.14% of them epithelium. HistoROI's probabilities overlapped heavily between normal and pathological. Normal-vs-abnormal ROC favoured the NBT-Classifier (DeLong, P < 0.0001). · Training exclusively on normal tissue enables…; Analysis of epithelium patches from BReAst…

4. SuggestedErrors cluster where tissue classes mix. Misclassified patches sit at the boundaries between feature clusters and have lower confidence. CAM and Grad-CAM highlight epithelial cell contents, collagen fibres and adipocyte membranes for the three classes.

   Of all external-validation patches, 98.2% had predicted-class probability > 0.7 and 84% > 0.99. On an example slide, patches below 0.99 lay mainly at compartment boundaries and inside large lobules. t-SNE of the last hidden layer gave well-separated clusters. On low-confidence patches, CAM and Grad-CAM attention was spread across mixed tissue. · NBT-Classifiers demonstrate strong…; Slide-level tissue compartment classification…; Visual assessment of feature representations; CAM and Grad-CAM visualisation

5. SuggestedAn end-to-end pipeline (HistoQC foreground detection, tiling into 128 or 256 µm patches, classification, connected-component lobule detection) outputs whole-slide class heatmaps and lobule plus peri-lobular masks that import into QuPath v0.3.0 for further analysis.

   Ran on slides with varying staining at 40x and 20x (Supplementary Fig. 21). After upscaling the epithelium map 32-fold, objects under 400,000 pixels are removed and holes under 400,000 pixels are filled (threshold 250,000 at 20x). Assessment is visual only; no lobule-detection accuracy is reported. · End-to-end pipeline for analysing large-scale…; Foreground tissue detection; Whole slide tissue classification; Lobule detection