josaklil-ai / surg-active-learning

Implementation for ALGES, an active learning method for label acquisition in semantic segmentation tasks.

ALGES: Active Learning with Gradient Embeddings for Semantic Segmentation of Laparoscopic Surgical Images

This is the official implementation for the paper Josiah Aklilu, Serena Yeung, "ALGES: Active Learning with Gradient Embeddings for Semantic Segmentation of Laparoscopic Surgical Images".

Run

To obtain the fully supervised model peformance on the held out test set for a particular dataset (e.g. 'cholecseg9k'), run 1 round of AL with the full size of the training set:

python run.py --n_init 4640 --n_rounds 1 --n_exp 1

To run the experiments from the paper, run the following for each dataset with a specific AL query strategy (e.g. 7 = ALGES-img or 8 = ALGES-seg):

For cholecSeg8k

• python run.py --dataset 'cholecseg8k' --test_size 1640 --n_init 50 --n_query 10 --n_rounds 30 --n_exp 3 --query 7

For m2caiSeg

• python run.py --dataset 'm2caiseg' --val_size 31 --n_init 10 --n_query 4 --n_rounds 51 --n_exp 3 --query 8

	Query method
1	Random
2	Max entropy sampling
3	Margins sampling
4	Least confidence sampling
5	Coreset
6	DEAL
7	ALGES-img (ours)
8	ALGES-seg (ours)

References

Some code adopted from the DeepAL and ViewAL repos.

1. Fu J, Liu J, Tian H, et al. Dual attention network for scene segmentation // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019: 3146-3154.
2. Xie S, Feng Z, Chen Y, et al. DEAL: Difficulty-aware Active Learning for Semantic Segmentation //

surg-active-learning's Issues

Can you tell me how to split the CholecSeg8k dataset? Which clips are included in the training set, test set, and verification set?

Can you tell me how to split the CholecSeg8k dataset? Which clips are included in the training set, test set, and verification set?

How did you preprocess the cholecseg8k masks

Hi there,

Very interesting work. I want to try running the code on the cholecseg8k dataset. Do you have any suggestions on how to preprocess the watershed masks? When I read them in NumPy arrays, the color code doesn't match with the dataset description. There are 17 unique color codes in the whole dataset, while there are only 13 classes.

Thanks