Table 3 Summary of top 10 ranked histologic features identified in the random forest model for the prediction of invasive cancer status among women with high and low % fibroglandular volume

From: Application of convolutional neural networks to breast biopsies to delineate tissue correlates of mammographic breast density

Feature Name

High global FGV (%) (> median)

Low global FGV (%) (≤ median)

High localized FGV (%) (> median)

Low localized FGV (%) (≤ median)

Global tissue amount

    

 Fat amount (µm2)

8

5

 Fat amount normalized (%)

3

7

 Stroma amount (µm2)

9

 Stroma amount normalized (%)

4

10

3

 Epithelium amount (µm2)

4

4

1

 Epithelium amount normalized (%)

6

7

8

Morphology

    

 Epithelial regions (IQ µm2)

1

4

 Epithelial regions (max µm2)

9

 Ecc epi regions (mean)

10

3

9

 Ecc epi regions (median)

2

2

 Ecc epi regions (IQ)

5

10

Spatial arrangement of the epithelial regions (Area-Voronoi diagram)

    

 Voronoi area (mean µm2)

5

7

6

 Voronoi area (median µm2)

3

5

 Voronoi area (SD µm2)

9

 Voronoi area (IQ µm2)

7

6

 Ratio epi to Voronoi (mean)

1

2

8

 Ratio epi to Voronoi (median)

2

1

 Ratio epi to Voronoi (IQ)

10

 Ratio epi to non-epi (median)

   

6

Spatial arrangement of the epithelial regions (Delaunay Triangulation)

    

 Neighbors (mean number)

8

  1. Ecc eccentricity, Epi epithelial, IQ interquartile, FGV fibroglandular volume, SD standard deviation
  2. Only histologic features ranked within the top 10 for prediction of each density measure are included in the table
  3. Features are ranked numerically and sequentially from 1–10, with 1 representing the most important feature and 10 representing the 10th most important feature
  4. The median cut points of breast density used in stratification were: global FGV (%) 34.4, localized FGV (%) 40.0