Design: Interobserver agreement study with structured training and re-evaluation
Population: 19 subspecialist breast pathologists scoring 54 images (cellularity, atypia) and 23 (border)
Primary outcome: Agreement on stromal cellularity, stromal atypia and border
Effect: Cellularity α 0.727 → 0.747; atypia α 0.570 → 0.554; border κ 0.532 → 0.615

Grading fibroepithelial lesions of the breast rests on stromal cellularity, stromal atypia and the tumour border. A multi-institution study asked 19 subspecialist breast pathologists to score 54 images for cellularity and atypia and 23 for border, then gave virtual training on a separate image set and repeated the exercise after at least three weeks.

Agreement on cellularity was moderate before and after training. Agreement on atypia was poor on a three-tier scale and only moderate when collapsed to two tiers (mild or moderate versus marked). Border agreement improved slightly but stayed moderate. Training mainly made pathologists more conservative, scoring the same cases lower.

The practical lesson is humility about the features that separate a cellular fibroadenoma from a benign or borderline phyllodes tumour. A two-tier approach is more reproducible than three, and borderline cases benefit from a second opinion and clear communication of uncertainty to the surgeon.

Use a two-tier assessment of stromal cellularity and atypia (mild or moderate versus marked) where possible; it was more reproducible.
Seek a second opinion on fibroepithelial lesions near a grading threshold.
State uncertainty in the report when the distinction affects surgical margins or follow-up.
Sample the tumour border generously; border assessment was only moderately reproducible even among experts.

Why it matters

Grade drives margin and follow-up decisions, yet the features behind it are less reproducible than reports often imply.

The statistics, in plain English

Kappa and alpha measure agreement beyond chance, from 0 (chance) to 1 (perfect). Values around 0.4 to 0.7, as here, mean pathologists agreed more than chance would predict but still disagreed often on the same image.