A labeling tool presents data and annotation controls for creating or reviewing machine-learning labels. Its design must match the label type, required expertise, time available, and product goal. section 6.2.4
Choose the tool and workflow around the required annotation:
- Binary classification asks whether an image contains malignancy. section 6.2.4
- Multiclass classification records which malignancy, if any, appears. section 6.2.4
- Detection identifies the region containing malignancy; segmentation marks its exact pixels. section 6.2.4
- Efficient operation also requires task decomposition, clear instructions, usable interfaces, quality control, aggregation, pricing, and process evaluation. section 6.2.4
Radiology images
A radiology team may label the same images with a binary, multiclass, detection, or segmentation interface. The choice changes labeling effort, team expertise, and the models and product the dataset can support. section 6.2.4
Figure 6.2 Various approaches to data labeling
Compare the annotation formats and the increasing detail they require. section 6.2.4
