## Self-Supervised Learning

## Self-Supervised Learning

Instance-level labels obtained using max-aggregation lack of large sensitivity, since max-aggregation only focus on the most distriminant instances. To alleviate this issue, training an Student model on distilled pseudolabels from the MIL-trained Teacher has shown promising results:

Instance-level labels obtained using max-aggregation lack of large sensitivity, since max-aggregation only focus on the most distriminant instances. To alleviate this issue, training an Student model on distilled pseudolabels from the MIL-trained Teacher has shown promising results:

```

```

python train_student.py --experiment_name instance_max

python train_student.py --experiment_name instance_max

```

```

**J. Silva-Rodríguez, A. Colomer, J. Dolz and V. Naranjo, "Self-Learning for Weakly Supervised Gleason Grading of Local Patterns," in IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 8, pp. 3094-3104, 2021.** [(paper)](https://ieeexplore.ieee.org/abstract/document/9361085)[(arXiv)](https://arxiv.org/abs/2105.10420)

**J. Silva-Rodríguez, A. Colomer, J. Dolz and V. Naranjo, "Self-Learning for Weakly Supervised Gleason Grading of Local Patterns," in IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 8, pp. 3094-3104, 2021.** [(paper)](https://ieeexplore.ieee.org/abstract/document/9361085)[(arXiv)](https://arxiv.org/abs/2105.10420)

## Proportion Constraints

## Proportion Constraints

Weakly supervised learning methods usually require large amounts od data to perform properly.In this project, we propose to introduce proportion priors per WSI as additional, weak information, to train realiable deep learning models. Concretely, we take advantadge of the Gleason scoring system, which defines a primary and secondary class per bag based on tissue proportions. Thus, we use inequality constraints to ensure that the percentage of positive predictions for the primary class is larger than the one for the secondary grade. Our formulation is flexible, and might deal with other problems, where more constraints are known regarding relative class proportions. You can train the proposed model using the following code.

Weakly supervised learning methods usually require large amounts od data to perform properly.In this project, we propose to introduce proportion priors per WSI as additional, weak information, to train realiable deep learning models. Concretely, we take advantadge of the Gleason scoring system, which defines a primary and secondary class per bag based on tissue proportions. Thus, we use inequality constraints to ensure that the percentage of positive predictions for the primary class is larger than the one for the secondary grade. Our formulation is flexible, and might deal with other problems, where more constraints are known regarding relative class proportions. You can train the proposed model using the following code.

```

```

python main.py --experiment_name instance_max_Constrained --aggregation max --mode instance --pMIL True --alpha_ce 1 --alpha_ic 0.1 --alpha_pc 1 --t_ic 15 --t_pc 5 --scheduler True --early_stopping True --criterion z --epochs 100

python main.py --experiment_name instance_max_Constrained --aggregation max --mode instance --pMIL True --alpha_ce 1 --alpha_ic 0.1 --alpha_pc 1 --t_ic 15 --t_pc 5 --scheduler True --early_stopping True --criterion z --epochs 100

```

```

Finally, you can produce visualization of instance-level predicitons through the following code:

Finally, you can produce visualization of instance-level predicitons through the following code:

```

```

python produce_visualizations.py --experiment_name instance_max_Constrained

python produce_visualizations.py --experiment_name instance_max_Constrained

```

```

## Contact

## Contact

For further questions or details, please directly reach out to Julio Silva-Rodríguez

For further questions or details, please directly reach out to Julio Silva-Rodríguez

(jusiro95@gmail.com)

(jusiro95@gmail.com)