Disagreement attention: Let us agree to disagree on computed tomography segmentation

Semantic segmentation is a popular technique successfully applied to various fields like self-driving cars, natural, medical, and satellite images, among others. On the one hand, a well-known concept of disagreement where two models help each other to learn better discriminative features comes from co-training. On the other hand, attention mechanisms are proven to improve the segmentation results; nevertheless, this technique solely focuses on signals with some kind of alignment. This research leverage both concepts in a new kind of attention based on disagreement (Pure, Embedded, and Mixed-Embedded disagreement attention) that improves the model generalisation. Furthermore, we introduce an innovative deep supervision approach (alternating deep supervision) which trains the model following the sequence of supervision branches. Extensive experiments over the segmentation benchmark datasets LiTS17 and CT-82 verify the effectiveness of the proposed approaches. The code is available at https://github.com/giussepi/disagreement-attention.

Automated summary

This research combines co-training and attention mechanisms to propose a new type of attention based on disagreement, which improves model generalization. It also introduces an innovative deep supervision approach that trains the model following the sequence of supervision branches. Extensive experiments on LiTS17 and CT-82 datasets verify the effectiveness of these approaches.