SNOW data in digital pathology - Supplementary materials

Adrien Foucart*, Olivier Debeir, Christine Decaestecker

LISA, ULB - CMMI, ULB

ShortRes Baseline architecture
Label Operation Kernel Size Output
Tile Size
Output
#Channels
Input     N 3
C1 Conv2D(Input) 3x3 N 64
R1_1 Conv2D(C1) 3x3 N 64
R1_2 Conv2D(R1_1) 3x3 N 64
R1_3 Conv2D(R1_2) 3x3 N 64
R1_A Add(C1, R1_3)   N 64
R1 MaxPool2D(R1_A) 2x2 N/2 64
R2_1 Conv2D(R1) 3x3 N/2 64
R2_2 Conv2D(R2_1) 3x3 N/2 64
R2_3 Conv2D(R2_2) 3x3 N/2 64
R2 Add(R1,R2_3)   N/2 64
R3_1 Conv2D(R2) 3x3 N/2 64
R3_2 Conv2D(R3_1) 3x3 N/2 64
R3_3 Conv2D(R3_2) 3x3 N/2 64
R3_A Add(R2,R3_3)   N/2 64
R3 MaxPool2D(R3_A) 2x2 N/4 64
U1 Conv2D_Transpose(R3) 2x2 N/2 64
R4_1 Conv2D(U1) 3x3 N/2 64
R4_2 Conv2D(R4_1) 3x3 N/2 64
R4_3 Conv2D(R4_2) 3x3 N/2 64
R4 Add(U1,R4_3)   N/2 64
U2 Conv2D_Transpose(R3) 2x2 N 64
R5_1 Conv2D(U2) 3x3 N 64
R5_2 Conv2D(R5_1) 3x3 N 64
R5_3 Conv2D(R5_2) 3x3 N 64
R5 Add(U2,R5_3)   N 64
C2 Conv2D(R5) 1x1 N 2
Output Softmax(C2)   N 2
ShortRes AutoEncoder architecture
Label Operation Kernel Size Output
Tile Size
Output
#Channels
Identical to ShortRes Baseline until R3
C2 Conv2D(R3) 3x3 N/4 16
R4_1 Conv2D(C2) 3x3 N/4 16
R4_2 Conv2D(R4_1) 3x3 N/4 16
R4_3 Conv2D(R4_2) 3x3 N/4 16
R4_A Add(C2,R4_3)   N/4 16
R4 MaxPool2D(R4_A) 2x2 N/8 16
U1 Conv2D_Transpose(R4) 2x2 N/4 16
U2 Conv2D_Transpose(U1) 2x2 N/2 16
Output Conv2D_Transpose(U2) 2x2 N 3
PAN Baseline architecture
Label Operation Kernel Size Output
Tile Size
Output
#Channels
Input     N 3
R1_1 Conv2D(Input) 3x3 N 64
R1_2 Conv2D(R1_1) 3x3 N 64
R1_S Conv2D(Input) 1x1 N 64
R1_A Add(R1_S, R1_2)   N 64
R1 MaxPool2D(R1_A) 2x2 N/2 64
R2_1 Conv2D(R1) 3x3 N/2 128
R2_2 Conv2D(R2_1) 3x3 N/2 128
R2_S Conv2D(R1) 1x1 N/2 128
R2_A Add(R2_S,R2_2)   N/2 128
R2 MaxPool2D(R2_A) 2x2 N/4 128
R3_1 Conv2D(R2) 3x3 N/4 256
R3_2 Conv2D(R3_1) 3x3 N/4 256
R3_S Conv2D(R2) 1x1 N/4 256
R3_A Add(R3_S,R3_2)   N/4 256
R3 MaxPool2D(R3_A) 2x2 N/8 256
R4_1 Conv2D(R3) 3x3 N/8 512
R4_2 Conv2D(R4_1) 3x3 N/8 512
R4_S Conv2D(R3) 1x1 N/8 512
R4 Add(R4_S,R4_2)   N/8 512
U1_1 Conv2D_Transpose(R4) 2x2 N/4 256
U1_2 Conv2D(U1_1) 3x3 N/4 256
U1_S Conv2D_Transpose(R4) 1x1 N/4 256
U1 Add(U1_S,U1_2)   N/4 256
U2_C Concat(U1,R2)   N/4 384
U2_1 Conv2D_Transpose(U2_C) 2x2 N/2 128
U2_2 Conv2D(U2_1) 3x3 N/2 128
U2_S Conv2D_Transpose(U2_C) 1x1 N/2 128
U2 Add(U2_S,U2_2)   N/2 128
U3_C Concat(U2,R1)   N/2 192
U3_1 Conv2D_Transpose(U3_C) 2x2 N 64
U3_2 Conv2D(U3_1) 3x3 N 64
U3_S Conv2D_Transpose(U3_C) 1x1 N 64
S1 Add(U3_S,U3_2)   N 64
S2 Resize(U2)   N 128
S3 Resize(U1)   N 256
F1 Conv2D(S1) 1x1 N 2
F2 Conv2D(S2) 1x1 N 2
F3 Conv2D(S3) 1x1 N 2
F_A Add(F1,F2,F3)   N 2
F Conv2D(F_A) 1x1 N 2
Output Softmax(F)   N 2
PAN AutoEncoder architecture
Label Operation Kernel Size Output
Tile Size
Output
#Channels
Identifcal to PAN Baseline until U1
S Resize(U1)   N 256
Output Conv2D(S) 1x1 N 3
Whole-Slide Results (PAN-GA50)

Segmentation results of the PAN-GA50 network (PAN using the Generated Annotations strategy with a 50% chance of using the output of the generator as annotation on negative patches during training). Normal tissue is colored in pink and artefacts are colored in green.

Block C
TCGA A1-A0SQ
TCGA AC-A2FB
TCGA AO-A0JE
TCGA D8-A141
The SNOW flowchart

Flowchart illustrating the process for analyzing a dataset through the SNOW framework.

Artefact dataset

To get the download link to the artefact dataset, please contact Adrien Foucart: afoucart@ulb.ac.be

Corrupted datasets

The corrupted noisy datasets were created by removing a certain percentage of the annotated objects from the supervision. As objects can vary in size, we verify on the corrupted GlaS dataset that we don't introduce a biais in some of the datasets by only removing small or big objects. The following graph shows the percentage of pixels removed from the annotations as a function of the percentage of objects removed.

Corrupted Dataset % objects remaining % pixels remaining
10% noise 89.47% 88.93%
20% noise 78.93% 81.84%
30% noise 69.18% 72.70%
40% noise 58.39% 61.90%
50% noise 47.33% 50.66%
60% noise 38.10% 42.34%
70% noise 28.48% 31.59%
80% noise 18.47% 22.08%
90% noise 11.05% 13.45%