Table 4 Comparison of Top 10 teams for IEEE Automatic Cancer Detection and Classification in Whole Slide Lung Histopathology Challenge 2019

From: Deep learning to assess microsatellite instability directly from histopathological whole slide images in endometrial cancer

 

Deep learning algorithm (Top 10 methods of 391 qualified international teams)

Performance evaluation

 

Team

Label Refine

Architecture

Preprocessing

Accuracy

Sensitivity

Specificity

Multi Model

PATECH

v

DenseNet & dilation block with Unet

Color normalization; Otsu to refine label

0.951

0.905

0.953

 

Byungje Lee

v

ResNet50 & DeepLab V3+

Multi data augmentations; Otsu to refine label

0.951

0.863

0.961

 

Turbolag

–

U-Net & ConvCRF

Multi-resolution training data

0.946

0.847

0.959

 

ArontierHYY

v

Mdrn80 +DenseNet & ResNet

Tile labeling strategy

0.929

0.856

0.940

 

Newhyun00

v

DenseNet103

Select clean labels

0.931

0.820

0.949

Single Model

CMIAS

v

DenseNet121 & FCN

Locate the tissue regions by a bounding box

0.938

0.800

0.957

 

Jorey

v

IncRes+ACF & CRF

Otsu to refine labels; Divided into 3 classes (tumor; normal; mix) and mix Mix file into other classes

0.937

0.815

0.951

 

Our pre-trained MFCN

–

FCN

None

0.923

0.860

0.928

 

Skyuser

–

ResNet18

Multi data augmentations;

0.932

0.767

0.956

 

Vahid

–

Small-FCN-521

None

0.921

0.846

0.932

  1. The top 3 methods are shown in bold format. Information reported in this table could be referred to Li et al.
Back to article page