Use Cases
MONAI-label has mainly three use cases:
- Cold-start annotation: The user can annotate a dataset from scratch using basic viewer capabilities (i.e. brushes, DeepGrow)
- Interactive label editing or modification: The user can modify or edit annotations created by an automatic model.
- Model quality improvement: This is designed to take advantage of an expert budget (i.e. 5 hours per week of a radiologist) to improve the quality of a segmentation model. For this, there are several active learning techniques and base learner models that allow this to happen.
The below table shows how the annotation time at the user's end can be reduced as more and more volumes are annotated. As the user trains on the data that they have annotated, the total time taken at user's end to annotate a 3D volume reduces. The Spleen dataset from medical segmentation decathlon was used for this experiment. A similar MONAI Label DeepGrow app can be found here.
The following analysis shows how the annotation time can be significantly reduced when using MONAI Label Apps based on DeepEdit paradigm. The assumptions/conditions made on this analysis are:
- We used the Heart MSD dataset (http://medicaldecathlon.com/). Four images for validation and 16 for training were used to obtain this validation plot.
- A skilled user takes 10 minutes to manually segment the left atrium in cardiac MR image
- One epoch using 2 images took aprox. 30 seconds. Adding an image makes the epoch take 10 more seconds
Assumptions/conditions made on this analysis are:
- We used the Spleen MSD dataset (http://medicaldecathlon.com/). Six images for validation and 35 for training were used to obtain this validation plot.
- A skilled user takes 10 minutes to manually segment the left atrium in cardiac MR image
- One epoch using 2 images took aprox. 50 seconds. Adding an image makes the epoch take 10 more seconds
