Learning to Segment From Scribbles Using Multi-Scale Adversarial Attention Gates.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Pixel-level image segmentation is challenging due to the difficulty of obtaining large, expert-annotated datasets, especially in medical imaging.
• Weakly-supervised learning offers a solution by utilizing less precise annotations like scribbles.

Purpose of the Study:

• To develop a novel weakly-supervised learning framework for image segmentation using scribble annotations.
• To leverage adversarial learning and attention mechanisms for improved segmentation accuracy and object localization.

Main Methods:

• A multi-scale Generative Adversarial Network (GAN) was trained with unpaired segmentation masks to generate realistic masks at multiple resolutions.
• Scribble annotations were used to guide the model in learning the correct object positions.
• A novel attention gating mechanism, conditioned with adversarial signals, was introduced as a shape prior for enhanced localization.

Main Results:

• The model achieved performance levels matching fully annotated methods across various medical and non-medical datasets.
• The attention gating mechanism, under adversarial conditioning, learned semantic attention maps, suppressed noise, and mitigated vanishing gradients.
• Extensions demonstrated success in semi-supervised learning, crowdsourcing scenarios, and multi-task learning.

Conclusions:

• The proposed method effectively utilizes scribble annotations for high-performance image segmentation.
• The adversarial attention gating mechanism provides a robust shape prior, improving multi-scale object localization.
• The framework is versatile and adaptable to various learning settings, including semi-supervised and multi-task learning.