Table 3 Summarizes the architecture design for the different findings.

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

Architecture

Finding

Block type

Depth

Width

Junction

Short path

C1

K1

Apply crop center on input (256 × 256)

Initial filter factor per path

Global pooling

No. of Ch. for 1st FC layer

Output function (no. of cat.)

Parameter capacity

Supervised contrastive learning53

Transformer48 (no. of blocks. header number)

Prostate Cancer

Inception

7

2

3

1

32

5 × 5

No

2

Max

96

Softmax (2)

178,342

No

No

Gleason pattern 3

Inception

4

2

3

1

32

5 × 5

No

2

Max

60

Softmax (2)

58,879

No

No

Inception

4

2

3

1

32

5 × 5

No

2

Max

60

Softmax (2)

58,879

No

No

Gleason pattern 4

Inception+

4

2

3

1

6

5 × 5

No

2

Avg

60

Sigmoid (1)

49,122

Yes

No

Inception+

4

2

3

1

6

5 × 5

No

2

Avg

60

Sigmoid (1)

49,122

Yes

No

Inception+

4

2

3

1

32

5 × 5

No

2

Avg

60

Softmax (2)

50,201

No

No

Soft attention

6

2

3

1

8

3 × 3

Yes

2

Max

20

Softmax (2)

51,732

No

No

Soft attention

6

2

3

1

8

3 × 3

Yes

2

Max

20

Softmax (2)

51,732

No

No

ResNet

5

2

3

1

16

5 × 5

No

4

Avg

44

Softmax (2)

206,638

No

No

 

Gleason pattern 5

Soft attention

5

2

3

1

16

5 × 5

No

2

Avg

24

Softmax (2)

48,590

No

No

Soft attention

5

2

3

1

16

5 × 5

No

2

Avg

24

Softmax (2)

48,590

No

No

Ductal morphology

Inception

4

2

3

1

16

5 × 5

No

2

Max

60

Softmax (2)

53,683

No

No

Cribriform pattern

Inception

3

2

3

1

32

5 × 5

No

2

Max

48

Softmax (2)

34,033

No

No

HGPIN

Inception

5

4

2

2

16

5 × 5

No

4

Max

160

Softmax (2)

412,322

No

No

Vessel

Inception+

5

2

3

1

16

5 × 5

No

2

Max

72

Softmax (2)

183,448

No

Yes (3,4)

Nerve

Inception+

5

2

3

1

16

5 × 5

No

2

Max

72

Softmax (2)

171,997

No

Yes (3,4)

Inflammatory cell infiltration

Inception+

5

2

3

1

16

5 × 5

No

2

Max

72

Softmax (2)

171,277

No

Yes (3,4)

Total

          

1012

 

1,878,587

  
  1. By combining all the models listed here, the novel architecture design achieved a total parameter capacity markedly lower than the parameter capacities of a single ResNet-18 model (~ 11 million) or the 2nd version of a single MobileNet model (~ 2.0 million parameters)50. These models combined a total of 1,012 features in the fully connected layers (in comparison, a single RestNet-18 model had 512 features in the fully connected layers). Given the compactness of our models, we could assign all 17 models to a single GPU card (VRAM 24 GB) for the pathology report generation task. For each Gleason pattern, we applied an ensemble model that weighted the predictions of the models equally. The weighted prediction scores ranged between 0 and 1. Note: The variation in the parameter capcity despite having the same hyperparameter configuration is due to the variation in the block number (depth) of the short path. A number of models with different configurations were tested to achieve the final model configuration via the grid search and trial-and-error approaches.
  2. Ch. Channel, conv. Convolution, C1 channel number of the first convolution layer, K1 kernel size of the first convolution layer, act. activation, no. number, cat. Category, FC fully connected; +  Layer normalization51 is applied instead of batch normalization52. Transfomer blocks were added prior to global pooling; since the output of the convolutional layer was batch size × height × width × channel, we reshaped the output to batch size × height ∗ width × channel before feeding into the transformer block; the output from the transformer48 was reshaped back to the dimension of the convolutional layer before global pooling was applied. For all the models, a patch dimension of 512 × 512 × 3 is applied.
  3. Significant values are in bold.
Back to article page