The authors propose an active learning strategy to train a model with a minimum amount of annotation. Once the network is trained with a part of the dataset, it is incrementally updated by learning from the most relevant annotations of the expert.

Paper review of the paper by Wei Shao, Liang Sun, Daoqiang Zhang accepted at IBSI 2018

Problem :
Annotation of histopathology images is a long and tedious process.

Idea :
Use an active learning strategy to train a model with a minimum amount of annotation. Once the network is trained with a part of the dataset, it is incrementally updated by learning from the most relevant annotations of the expert. The model is then retrained with the new annotations given by the expert, with the goal of reducing the model’s uncertainty. It has the advantage of lowering the annotation time and the redundancy in the images (some images may not be useful to the model because they are very similar, thus do not bring new information to the system).

Solution :
Classical patch based CNN segmentation model learned with cross-entropy and a constraint term. The constraint is here to preserve intra/inter class distribution, simply by minimizing the Euclidean distances of features from the same class and maximizing distances of features from different classes. The constraint is called pairwise constrained regularizer. The objective is formulated as : $loss = cross\ entropy(cnn(x), y) + \alpha \times constraint(cnn(x))$

Once the base model is trained, the patches with the highest cross-entropy scores are chosen from the unannotated dataset, the expert is asked to segment them because they are what the model is having serious difficulties to classify.

Results :
By varying the size of the training dataset, the F-score between a classical deep learning model and the proposed solution are compared. We see that the proposed model is able to reach the best F-score before the classical approach.

Personal Opinion :
It’s not clear to me if the improvement is because of the training strategy or the constraint term. At least I am not able to quantify the contribution of each. Globally speaking, I think this type of training strategy would be useful in a practical setup where the annotation process is critical.

The authors are encouraged to provide any information that could help to understand this work better, or inform about any followup of it.