A segmentation mask is an image-like map that labels each pixel according to whether it belongs to a target region. In binary semantic segmentation, true pixels identify nodule candidates and false pixels identify uninteresting healthy tissue. section 15.2
Masks pair with the original CT slices to form a training dataset. They provide ground-truth regions for fine-tuning a model such as SegFormer, while avoiding the cost of manually annotating every pixel. section 15.6.3
A practical mask dataset includes:
- A CT image and its corresponding mask, stored with matching filenames or metadata. section 15.6.3
- A binary representation in which true values mark the target region and false values mark background tissue. section 15.2
Generating a mask with SAM
A point placed at a suspected nodule center can prompt SamPredictor to produce a mask. The resulting mask is converted to an image and saved alongside the CT slice for later fine-tuning.
