We discussed how to preprocess 3D volumes for tumor segmentation in the previous article, so in this article we will discuss another important step when working on a deep learning project. This is the data augmentation step.
We are all aware that in order to train a neural network, a significant amount of data is required in order to obtain an accurate model as well as a robust model that can work with the majority of cases in that specific task. However, it is not always possible to obtain a large amount of natural data in any task, particularly in healthcare projects. Because one input in medical imaging is a single patient with multiple slices, and we all know how difficult it is to assemble a dataset of this type of data (a lot of patients).
For this reason, we must assist ourselves by creating synthetic data in order to improve our results slightly.
The generation of these synthetic data is referred to as data augmentation, which means that we start with our dataset and then perform some transformation to generate new data.
If you have previously done data augmentation for a project with a normal task and in 2D images, you will understand what I am trying to say; otherwise, don’t worry, we will take it one step at a time until we get the results.
To generate these synthetic data, we must first apply some affine transforms to the original data. These transformations can include rotations, zooming, translations (shifting), noises, flipping, and so on.
But be careful, because while working with normal images, we can use any of these transforms; however, when working with medical images, we cannot use all transforms because it may create a shape that has no relation to the human body, which is not the point.
And since we’ll be working with 3D volumes, the transformations will be even worse if we’re not careful.
We will always use the same monai that we used for processing for this operation. For those who are unfamiliar with monai, it is an open source framework based on Pytorch that can be used to segment or classify medical images.
In my experience, there are only a few transforms that can be mixed together at random to produce synthetic patients. Here are the best transformations that I found to be effective:
- Shifting (translation)
- Gaussian noise
So I’ll show you how to use monai to easily apply these transformations.
There is something you should know about how to apply these transforms. As I previously stated, monai is based on Python, so…
Continue reading: https://towardsdatascience.com/3d-volumes-augmentation-for-tumor-segmentation-using-monai-1b6d92b34813?source=rss—-7f60cf5620c9—4