Feature-wise normalization for 3D medical images - Revision history

Amira at 13:40, 29 September 2020

2020-09-29T13:40:00Z

Amira at 13:38, 29 September 2020

2020-09-29T13:38:10Z

Amira: Amira moved page Name of the new project to Feature-wise normalization for 3D medical images

2020-09-29T13:35:30Z

Amira moved page Name of the new project to Feature-wise normalization for 3D medical images

Amira at 13:29, 29 September 2020

2020-09-29T13:29:49Z

Tiago: Created page with "{{StudentProjectTemplate |Summary=The topic focuses on generative models (GAN) for CAN-bus data and investigating the representation learning capabilities of such techniques ..."

2020-09-29T07:28:45Z

Created page with "{{StudentProjectTemplate |Summary=The topic focuses on generative models (GAN) for CAN-bus data and investigating the representation learning capabilities of such techniques ..."

New page

{{StudentProjectTemplate
|Summary=The topic focuses on generative models (GAN) for CAN-bus data and investigating the representation learning capabilities of such techniques
|Keywords=GAN, CAN data, MAR
|TimeFrame=2020 Fall - 2021 Summer
|References=https://papers.nips.cc/paper/8789-time-series-generative-adversarial-networks.pdf

https://arxiv.org/abs/1706.02633

https://openreview.net/pdf?id=rJedV3R5tm

https://www.aaai.org/Conferences/AAAI/2017/PreliminaryPapers/12-Yu-L-14344.pdf

https://arxiv.org/pdf/1511.06434.pdf

|Prerequisites=Excellent Programming Skills
Excellent knowledge in Machine Learning and Neural Networks
|Supervisor=Kunru Chen, Tiago Cortinhal, Thorsteinn Rögnvaldsson,
|Level=Master
|Status=Internal Draft
}}
Control Area Network (CAN) is a protocol that is used to manipulate vehicles. It is multidimensional and consists of control and sensor signals to and from different parts of the equipment. Since this data comes internally from the machine itself, it is stable and cheap to collect it. Previous work has shown that CAN data can be used to build representations for machine activity recognition (MAR) for forklift trucks. However, those representations are limited to only describing the existing data in both realism and diversity. Creating representation by training a vanilla autoencoder has disadvantages when trying to explore the entire space of CAN signals.

Generative approaches have been used mostly in traditional types of data, like images, and have shown to have great capabilities to learn the underlying distribution as well as allowing us to sample new unseen data points. This has shown great results as we can see in https://thispersondoesnotexist.com, or even in pictures to picture translations and style transfers. This generative capability also allows us to perform arithmetic operations on the vector and see the underlying structure of each different “class” of outputs.

Nevertheless, the work done in other data modalities is still sparse but nevertheless growing in interest. In this thesis, the main interest is focused on a very specific type of data that might bring all kinds of hardships and obstacles to overcome. Some of those hardships might come from the type of data we are trying to generate. This needs to be investigated and solutions to overcome these types of situations are a key aspect we will be looking for.
The students need to develop a GAN-based network to generate CAN data, to evaluate the quality of the generated data, and to use that data in a MAR task.

Research Questions:
Can GANs generate realistic CAN data?
Can GANs generate/predict the (near) future CAN signals?
Is the latent space an informative representation about the CAN signals?

@@ Line 2: / Line 2: @@
 |Summary=Normalization of 3D medical imaging either as a data pre-processing or as feature-wise batch normalization during CNN model training
 |Keywords=CNN, 3D models
 |Supervisor=Amira Soliman, Stefan Byttner,  Kobra Etminani
 |Level=Master

@@ Line 1: / Line 1: @@
 {{StudentProjectTemplate
-|Summary=Normalization of 3D medical imaging either as a data reprocessing or as feature-wise batch normalization during CNN model training
+|Summary=Normalization of 3D medical imaging either as a data pre-processing or as feature-wise batch normalization during CNN model training
 |Keywords=CNN, 3D models
 |TimeFrame=2020 Fall - 2021 Summer

← Older revision	Revision as of 13:35, 29 September 2020
(No difference)

@@ Line 1: / Line 1: @@
 {{StudentProjectTemplate
-|Summary=The topic focuses on generative models (GAN) for CAN-bus data and investigating the representation learning capabilities of such techniques
+|Summary=Normalization of 3D medical imaging either as a data reprocessing or as feature-wise batch normalization during CNN model training
-|Keywords=GAN, CAN data, MAR
+|Keywords=CNN, 3D models
 |TimeFrame=2020 Fall - 2021 Summer
 |Prerequisites=Excellent Programming Skills
 Excellent knowledge in Machine Learning and Neural Networks
-|Supervisor=Kunru Chen, Tiago Cortinhal, Thorsteinn Rögnvaldsson,
+|Supervisor=Amira Soliman, Stefan Byttner,  Kobra Etminani
 |Level=Master
-|Status=Internal Draft
+|Status=Open
 }}
-Control Area Network (CAN) is a protocol that is used to manipulate vehicles. It is multidimensional and consists of control and sensor signals to and from different parts of the equipment. Since this data comes internally from the machine itself, it is stable and cheap to collect it. Previous work has shown that CAN data can be used to build representations for machine activity recognition (MAR) for forklift trucks. However, those representations are limited to only describing the existing data in both realism and diversity. Creating representation by training a vanilla autoencoder has disadvantages when trying to explore the entire space of CAN signals.
+Normalization is a required preprocessing step, especially for deep learning and convolutional neural networks, such that the network becomes unbiased towards the different features. However, in medical images, the whole intensity normalization may lead to reduced sensitivity for relatively important features. The objective of this master thesis is to study the state-of-the-art normalization techniques used in 2D images, investigate the applicability of such techniques in 3D medical images, and apply them either as a preprocessing step or as feature-wise batch normalization during the model training.