In CT image processing, the segmentation model usually serves as the foundation part for a wide range of applications. To train a segmentation model, an essential part is data annotation.
Manual annotation is still unavoidable at this stage since the unsupervised learning methods still can't serve the tasks very well. MITK is quite suitable for data annotation. A very useful tool. The annotation data should be transferred into the format of tiff or nifti like what we talk about in part 1.
Besides manual annotation, another useful technique is semi-supervised learning which is quite efficient for generating a good model with few annotation data. A typical way to avoid a large number of manual annotations is employing the pseudo-label:
- Annotate several data manually;
- Generate a semi-supervised/supervised model and retrain the model;
- Apply the model on all collected, and select good cases to retrain the model.
- Redo 2~3 until enough data have been collected.
After the data annotation, data verification is a must. Every annotation should be carefully reviewed to ensure the training data are correct. Here, a good way is to generate overlapping images for quick verification.
标签：good,processing,verification,model,data,annotation From： https://www.cnblogs.com/xiaoxu-xli/p/16964343.html