what are the important feature of a biomedical image for classification.

Question

Md. Rokibul Islam 2016-11-29

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/314484-what-are-the-important-feature-of-a-biomedical-image-for-classification

评论： Image Analyst 2016-12-1

采纳的回答： Walter Roberson

i want to work with NN for image classification. so i need to extract some important feature

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Walter Roberson 2016-11-29

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/314484-what-are-the-important-feature-of-a-biomedical-image-for-classification#answer_245162

There is only three essential properties of biomedical image for the purpose of classification by computers:

The image must have the property of existence. Non-existent images cannot be classified by computer
The image must have the property of reality, in the sense of not existing only in fiction. For example, Spiderman exists in fiction and so satisfies the property of existence, but Spiderman does not have the property of reality.
The image must be represented in an electronic format that can be used with the computer doing the classification. For example it is fine if the image exists only in trinary (base 3 representation, three-level representation) provided that you are using a computer that can read trinary.

There are no other features that are required for classification of all biomedical images. Every mathematical or positional property of biomedical images is useful for classifying some biomedical images and irrelevant for classifying other images. Furthermore, for any given image, there might be different features that are useful for classification of different aspects. Even the parts you might think of as complete "background" might turn out to be key to extracting something completely different from what you were originally looking for.

5 个评论
显示 3更早的评论隐藏 3更早的评论

Walter Roberson 2016-12-1

There is a lovely quote about the varieties of non-existent beings that often appeared on Usenet in people's .sig but I do not seem to be able to find it at the moment.

It turns out that there is a whole bodies of serious study about varieties of non-existence; https://en.wikipedia.org/wiki/Meinong's_jungle and http://plato.stanford.edu/entries/nonexistent-objects/

For example the logically impossible (such as square circles) are somehow different from the purely imaginary (such as unicorns); and those are somehow different again from the things that could reasonably exist but do not happen to (e.g., there is no stuffed animal on my desk right now as I write this, but there has been in the past and I could easily put one there.)

Image Analyst 2016-12-1

Well unicorns must exist. Otherwise how could Amazon.com sell these:

请先登录，再进行评论。

Answer 2

Image Analyst 2016-11-29

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/314484-what-are-the-important-feature-of-a-biomedical-image-for-classification#answer_245243

Google "convolutional neural networks".

These are useful where you don't know exactly what feature you want to measure so you just train the system with thousands of images of known classes and it applies simple things like multi scale convolutions (say, to find edges) and adjusts the weights to identify patterns in the image. Then during application of a test image it will apply the model/weights/CNN to make a guess as to what class your test image belongs.

Here is the course on deep learning that includes convolutional neural networks. One of my coworkers has worked through a number of courses on this topic and found this course to far exceed the rest. https://www.udacity.com/course/deep-learning--ud730

There is a package within Matlab to do convolutional neural networks. http://www.mathworks.com/help/nnet/convolutional-neural-networks.html

4 个评论
显示 2更早的评论隐藏 2更早的评论

Walter Roberson 2016-11-30

https://www.mathworks.com/help/nnet/convolutional-neural-networks.html

"NOTE: Training a convolutional neural network requires the Parallel Computing Toolbox™ and a CUDA®-enabled NVIDIA® GPU with compute capability 3.0 or higher."

Image Analyst 2016-11-30

OK, for the MATLAB solution, yes.

Actually the guy I work with is not using MATLAB. He said "Because these models typically train faster with NVIDIA GPUs (using the CUDA toolkit), I normally use the Torch language http://torch.ch/ for Deep Learning on a linux server. There are also free Deep Learning packages in the Python language."

请先登录，再进行评论。