Table 2 The description of the cohort utilized to run the test conditions (external validation set).

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

Test condition

Data description or project source (number of cases), image size

Number of images/patches

Scanner vendor

Objective magnification level

Prostate cancer detection on the H&E slides

Slide condition or sample dimension

Old slides (~ 20 years)

Patch images

McNeal’s anatomy study32

(15 cases), ~ 512 × 512 µm

11,862+

Philips

10×

Recent slides (< 6 years)

TMA Spots (Smallest sample dimension)

Stanford TMA database

(339 cases), 2048 × 2048 pixels spot images

1129

Leica

20×

All whole mount slides of each single case+++ (Largest sample dimension)

Radical prostatectomy

(46 cases), whole-mount slide

368

Leica

20×

Tumor volume estimation

All slides of each single case+++

Radical prostatectomy (46 cases), whole-mount slides

368

Leica

20×

Sorting slides according to prostate cancer presence

Tissue sampling method

Radical prostatectomy specimens

Radical prostatectomy

(136), whole-mount slides

1080

Leica

20×

Dissected pelvic lymph nodes

Lymph node dissection

(50 lymph nodes), whole slides

19

Leica

20×

Obduction

Cystprostatectomy specimens

The Genotype-Tissue Expression project33 (40 cases), whole slides++

40

Leica

40×

Radical prostatectomy

specimens

PAI-WSIT project34 (18 cases), whole slides++

60

Hamamatsu

40×

Gleason pattern detection and ISUP grading

Tissue sampling method

Biopsy cores

The International Society of Urological Pathology image library35, 2048 × 2048 pixels (72dpi) micrograph, (137 cases)

137

Olympus (micrograph)

20×

Radical prostatectomy specimens

Radical prostatectomy (136), whole-mount slides

1080

Leica

20×

Very limited tissues with prostate cancer

Patches from 594 random regions with Gleason patterns 3–5 and HGPIN in 24 whole-mount WS images (24 cases), 512 × 512 µm

3840 (20×)

1128 (10×)

Leica

20×

10×

Detection of ductal morphology

Patch images

Patches from 38 random regions from 2 WS images of 2 cases with ductal adenocarcinoma plus 218 random regions with Gleason pattern 3–5 in 9 WS images of 9 cases, 512 × 512 µm

2112

Leica

10×

Detection of cribriform pattern

Patch images

Patches from 32 random regions with Cribriform patterns of 5 cases and 199 random regions with noncribriform prostate cancers and Gleason patterns 3–5 in 9 whole-mount WS images (9 cases), ~ 512 × 512 µm

928

Leica

10×

Detection of vessels

Patch images

Patches from 642 random regions with blood vessels on 22 WS images (22 cases) and 478 random regions with Gleason patterns 3–5 on 20 WS images (20 cases), 512 × 512 µm

4608

Leica

10×

Detection of nerve structure

Patch images

Patches from 628 random regions with nerves or ganglions on 22 WS images (22 cases) and 216 random regions with Gleason patterns 4–5 (8 slides, 8 cases), 512 × 512 µm

1280

Leica

10×

Detection of inflammatory cell infiltration

Patch images

Patches from 123 random regions with inflammatory cell infiltration on 19 WS images (19 cases) and 216 random regions with Gleason patterns 4 and 5 (8 slides, 8 cases), ~ 512 × 512 µm

768

Leica

10×

HGPIN detection

Patch images

Random 32 regions from 10 WM images (10 cases); 40 random regions with intraductal adenocarcinoma from 4 WM images (4 cases); 19 random regions with benign prostatic hyperplasia from 4 WM images (4 cases), ~ 512 × 512 µm

2687

Leica

10×

Integration into an electronic pathology report platform

Radical prostatectomy specimens

136 radical prostatectomy specimens, complete representative whole-mount slides per case

1028

(Median: 8 per case)

Leica

20×

  1. All histological slides were stained with hematoxylin and eosin. HGPIN: high-grade prostatic intraepithelial neoplasia. + the whole-slide images were tiled into small image patches ++whole slides with a portion of the prostatic slice (2.3 times smaller than the prostatic slice). TMA: tissue microarray. +++Given that a single whole-mount (WM) slice roughly corresponds to 20–30 biopsy cores and because of the time- and labor-intensive effort required for high-precision annotation of WM images for prostate cancer, we randomly selected 46 cases with a total of 368 WM images (~ 7,360–11,040 biopsy core images or ~ 894,240 nonoverlapping tiles (dimensions: 512 × 512 pixels) at 10 × and ~ 1 µm per pixel for patch-level performance evaluation. * The images were obtained at different sites. The negative control groups for the detection of ductal morphology, the cribriform pattern, vessels, nerve structure, and HGPIN are described in the sections below.