Discrimination is a model’s ability to separate two classes. For nodule detection, it means distinguishing actual nodules from normal anatomical structures—the central goal of training. section 14.4.1
Training data should give the model a fair chance to learn that separation:
- Balance positive and negative samples instead of presenting long runs of negatives. section 14.4.1
- Mix both labels within batches so their competing signals can guide learning. section 14.4.1
- Repeat scarce positive samples when the dataset contains only a small fixed number of them. section 14.4.1
- Treat dataset bias as a separate risk: uncorrected real-world biases can persist in the trained model. section 14.4.1
Balanced LUNA training
A training application can expose a --balanced option, pass a ratio setting to LunaDataset, and alternate positive and negative candidates while reusing positive samples as needed.
