01.12.2018  Research article  Ausgabe 1/2018 Open Access
Classification of lung nodules in CT scans using threedimensional deep convolutional neural networks with a checkpoint ensemble method
 Zeitschrift:
 BMC Medical Imaging > Ausgabe 1/2018
Background
Method
Layer connection
Model description
Layer name

Structure


convolution_1

7×7×7 conv3×3×3 max pool

convolution_2

\(\begin {bmatrix} 3 \times 3 \times 3 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×2

convolution_3

\(\begin {bmatrix} 3 \times 3 \times 3 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×2

convolution_4

\(\begin {bmatrix} 3 \times 3 \times 3 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×2

convolution_5

\(\begin {bmatrix} 3 \times 3 \times 3 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×2

7×7×7 avg pool1000d FCsoftmax

Layer name

Structure


7×7×7 conv


3×3×3 max pool


Dense block

\(\begin {bmatrix} 1 \times 1 \times 1 \text { conv}\\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×6

Transition

1×1×1 conv2×2×2 avg pool

Dense block

\(\begin {bmatrix} 1 \times 1 \times 1 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×12

Transition

1×1×1 conv2×2×2 avg pool

Dense block

\(\begin {bmatrix} 1 \times 1 \times 1 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×24

Transition

1×1×1 conv2×2×2 avg pool

Dense block

\(\begin {bmatrix} 1 \times 1 \times 1 \text { conv} \\ 3 \times 3 \times 3 \text { conv} \end {bmatrix}\) ×16

7×7×7 avg pool1000d FCsoftmax

Ensemble
Experiment and result
Dataset
Preprocessing
Evaluation metric
Result
Setup name

Model type

Input size

# of checkpoints

Ensemble


S48

3D shortcut DCNN

48

1

X

S64

3D shortcut DCNN

64

1

X

D48

3D dense DCNN

48

1

X

D64

3D dense DCNN

64

1

X

ESBS48

3D shortcut DCNN

48

6

O

ESBS64

3D shortcut DCNN

64

6

O

ESBS

3D shortcut DCNN

48

6

O

64

6


ESBD48

3D dense DCNN

48

6

O

ESBD64

3D dense DCNN

64

6

O

ESBD

3D dense DCNN

48

6

O

64

6


ESBBEST

3D shortcut DCNN

48

1

O

64

1


3D dense DCNN

48

1


64

1


ESBALL

3D shortcut DCNN

48

6

O

64

6


3D dense DCNN

48

6


64

6

0.125

0.25

0.5

1

2

4

8

CPM



S48

0.691

0.788

0.851

0.891

0.910

0.934

0.945

0.859

S64

0.736

0.818

0.880

0.911

0.932

0.950

0.960

0.884

D48

0.676

0.765

0.839

0.894

0.922

0.938

0.953

0.855

D64

0.710

0.800

0.870

0.902

0.924

0.943

0.958

0.872

ESBS48

0.655

0.739

0.863

0.927

0.962

0.973

0.976

0.871

ESBS64

0.633

0.744

0.870

0.943

0.974

0.980

0.980

0.875

ESBS

0.683

0.813

0.911

0.954

0.969

0.982

0.982

0.899

ESBD48

0.645

0.736

0.816

0.908

0.954

0.975

0.980

0.859

ESBD64

0.646

0.736

0.834

0.919

0.962

0.977

0.981

0.865

ESBD

0.679

0.778

0.878

0.937

0.963

0.981

0.981

0.885

ESBBEST

0.734

0.814

0.895

0.934

0.957

0.971

0.976

0.897

ESBALL

0.720

0.842

0.914

0.954

0.974

0.982

0.982

0.910

Predicted class


D48

Nodule

Nonnodule


Actual

Nodule

0.913

0.087

Class

Nonnodule

0.016

0.984

Predicted class


EBSALL

Nodule

Nonnodule


Actual

Nodule

0.933

0.067

Class

Nonnodule

0.007

0.993

Method

0.125

0.25

0.5

1

2

4

8

CPM



LUNA16CAD

2D CNN

0.113

0.165

0.265

0.465

0.596

0.695

0.785

0.440

LungNess

2D CNN

0.453

0.535

0.591

0.635

0.696

0.741

0.797

0.635

iitem03

2D CNN

0.394

0.491

0.570

0.660

0.732

0.795

0.851

0.642

[
22]

3D CNN

0.517

0.602

0.720

0.788

0.822

0.839

0.856

0.735

LUNA16CAD

3D CNN

0.640

0.698

0.750

0.804

0.847

0.874

0.897

0.787

[
9]

2D CNN

0.734

0.744

0.763

0.796

0.824

0.832

0.834

0.790

DIAG_CONVNET [
23]

3D CNN

0.636

0.727

0.792

0.844

0.876

0.905

0.916

0.814

UACNN

2D CNN

0.655

0.745

0.807

0.849

0.880

0.907

0.925

0.824

CUMedVis [
24]

3D CNN

0.677

0.737

0.815

0.848

0.879

0.907

0.922

0.827

D48

3D CNN

0.676

0.765

0.839

0.894

0.922

0.938

0.953

0.855

ESBALL

3D CNN

0.720

0.842

0.914

0.954

0.974

0.982

0.982

0.910
