data augmentation in CNN

Hello

I am going to train a squeezenet to detect CT scans in two classes of people with covid19 and healthy people, assuming that each image is 512 x 512.

At least in the training phase, how many images should be present so that there is no need to data augmentation like rotation and addition of gaussian and.....

Answer 1

[Walter Roberson]

At least 83886080 images for training under the circumstances you describe.

... This should suggest to you that data augmentation is a very important stage.

[lech king]

Wow

I understand what you mean

Thanks for your help

I only had a few questions

I am new to this subject and I imagined that for example the image rotation will cause the network to not have a proper training or when adding noise to the image

Isn't that so?

[Walter Roberson]

It depends on the processing. There are features that are largely independent of orientation of the image, and if you are able to use only those kind of features then the rotation of the image will matter (less).

However, it is more common for features to depend upon rotation, and partly upon scale. If you do not have a rotation layer for the training data, then the network will be trained to recognize only items that are very close to the same orientation as the data you trained on. If you do have a rotation layer, then your system should be able to withstand up to 45 degree rotation (and possibly more) for the input images.

Patients are rarely exactly where they should be ideally, and rarely rotated exactly as you would like, so unless you have someone going through the images and fixing them up by hand before they make it into your training or prediction system, then you could have problems.

[lech king]

Thank you very much

Because I work on CT scans of covid19, which are lung images

In about 70% of cases, the symptoms of the disease appear in certain places

But the reference article I am working on, because it has 441 images in training phase , has thus strengthened its database and achieved 85% accuracy.

a rotation (with a random angle between 0 and 90 degrees), a scale (with a random

value between 1.1 and 1.3) and addition of gaussian noise

to the original image

[lech king]

I use MATLAB Deep Learning application. In this section, rotation and scale can be easily selected.

Thanks for the tips on how to add noise to the original images MATLAB Deep Learning application