Indian Institute of Technology Kharagpur

Department of Electrical Engineering | Signal Processing and Machine

Learning

U-Net : Convolutional Network for

Biomedical Image Segmentation

G B Kevin Arjun
MTech – Signal Processing and Machine Learning
23EE65R19

Under the supervision of

Dr. Debdoot Sheet
 Indian Institute of Technology Kharagpur
Department of Electrical Engineering | Signal Processing and Machine
Learning

Outline
• Introduction and Motivation

• Architecture

• Data Augmentation

• Training

• Conclusion

2
 Indian Institute of Technology Kharagpur
Department of Electrical Engineering | Signal Processing and Machine
Learning

Introduction and Motivation

• Deep Convolutional Networks have outperformed the state of the art in visual recognition

• Amount of training data and size of considered network

• Biomedical image processing applications require semantic segmentation and localisation

• An extended fully convolutional network working with very few training images

• Designed architecture mainly consist of contraction path , expansion path , bridge

3
 Indian Institute of Technology Kharagpur
Department of Electrical Engineering | Signal Processing and Machine
Learning

Architecture
1 64 64 64 64 2
128

Skip Connections Output

Input
image segmentation
tile map

Expansion Path
570 x 570

570 x 570
570 x 570

570 x 570
Contraction Path

572 x 572

570 x 570

570 x 570

128 128 256 128

2
2

196
200

198
2

282 2

2
284

280

256 256 256

Conv 3x3,ReLU
512

Up-conv 2 x 2
2

138 2

136 2
140

2
2

2
104

102

100
Copy and crop
1024 512
56 2
64 2

2
54 2
68 2

66 2

52
1024 Max pool 2x2
Conv 1x1
32 2

30 2

28 2

4
Bridge
 Indian Institute of Technology Kharagpur
Department of Electrical Engineering | Signal Processing and Machine
Learning

Data Augmentation
• Supplements the creation of data variations

• Elastic Deformation for data augmentation during training

5
 Indian Institute of Technology Kharagpur
Department of Electrical Engineering | Signal Processing and Machine
Learning

Training
• Stochastic Gradient Descent for segmentation maps
• High momentum (0.99) based gradient descent
• Energy function is computed by pixel wise soft-max and cross Entropy loss function
• Small separation border is computed using morphological operations

6
 Indian Institute of Technology Kharagpur
Department of Electrical Engineering | Signal Processing and Machine
Learning

Conclusion
• A good tradeoff between context information and localization

• Computationally efficient

• Trainable with a small data-set

• Trained end-to-end

• Preferable for bio-medical applications