SISSI: Seamless Iterative Semi-Supervised Correction of Imperfect Labels in Microscopy Images

Introduction

This is the official PyTorch implementation of the paper "SISSI: Seamless Iterative Semi-Supervised Correction of Imperfect Labels in Microscopy Images" to appear in MICCAI 2022 Workshop on Domain Adaptation and Representation Transfer DART 2022.

Noisy annotiation generation

The code for noisy annotation generation is in noisy_annotations_generation. Specific algorithms have been developed for different state of cells: dead, alive and inhibited, the noisy image level annotations are assumed to be true when developing these algorithms.

Training your model

You can run deep_learning_code/train_mix.py to train your model in SSSI framework.

SISSI Components

Determining the Start of the Semi-Supervised Phase

• ADELE Adoption for Object detection
• The implementation for determining the optimal point that represents the start of memorisation phase can be found in if_update in deep_learning_code/utils.py.

Pseudo Label Generation

• We adopt the implementation of TTA and Weighted Boxes Fusion from kentaroy47 for pseudo label generation.
• The code can be found in deep_learning_code/odach.

Synthetic-like image adaptation according to pseudo labels (Seamless cloning)

• The use of seamless clone can be found in seam_less_clone in deep_learning_code/dataloaders/utils.

• The iterative update of synthetic-like images is handled by cell_lab_dataset in deep_learning_code/dataloaders/instance_seg_dataset.py and the flag to perform this update is controlled in correct_labels in deep_learning_code/reference/engine.py.

Citation

If this code is useful for your research, please consider citing:

@article{Elbatel2022,
   author = {Marawan Elbatel and Christina Bornberg and Manasi Kattel and Enrique Almar and Claudio Marrocco and Alessandro Bria},
   doi = {10.48550/arxiv.2208.03327},
   month = {8},
   title = {Seamless Iterative Semi-Supervised Correction of Imperfect Labels in Microscopy Images},
   publisher = {arXiv},
   url = {https://arxiv.org/abs/2208.03327v1},
   year = {2022},
}