Table 8 The optimization set for Gleason pattern detection used to select the best models according to the best area under the curve (AUROC).

From: Critical evaluation of artificial intelligence as a digital twin of pathologists for prostate cancer pathology

Finding

Case, n (%)

Regions, n (%)

Patches, n (%)

GP3

12 (21.4)

32 (6.0)

923 (11.1)

GP4

9 (16.1)

166 (31.0)

1318 (15.9)

GP5

20 (35.8)

216 (40.2)

5397 (65.0)

Normal tissue

11 (19.6)

106 (19.8)

466 (5.6)

HGPIN

4 (7.1)

16 (3.0)

204 (2.4)

  1. This optimization set was curated from The Cancer Genome Atlas images with demarcation of 536 prostate cancer heterogeneous regions representing different Gleason patterns and benign tissues in 35 patients (~ 15 regions per patient) by a team of a senior pathologist and a prostate cancer researcher. These patches (512 × 512 pixels) were curated at the 10 × magnification level, resulting in a total of 8308 patches. The “Normal tissue” category covers the prostatic epithelium, stroma, and atrophy. HGPIN stands for high-grade prostatic intraepithelial neoplasia and is the presumed precursor of prostate cancer.
Back to article page