From Explanations to Segmentation: Using Explainable AI for Image Segmentation

Reduces labeling burden: The paper demonstrates that explainable AI techniques can enable image segmentation with only image-level labels rather than requiring pixel-by-pixel manual annotation, significantly lowering the cost and effort of preparing training data for real-world applications.

Achieves comparable performance with less effort: The proposed method using Layer-wise Relevance Propagation (LRP) produces segmentation results similar to established architectures like U-Net, but requires substantially simplified training data generation, making it more practical for deployment.

Makes AI more accessible: By reducing the need for expensive pixel-precision labeling, this approach democratizes access to segmentation technology for organizations and domains where detailed manual annotation is impractical or prohibitively costly.

Transparency serves a practical purpose: The paper illustrates how explainability methods—traditionally used to understand AI decisions—can be repurposed as a tool to reduce annotation requirements, showing that interpretability and efficiency can work together rather than being competing concerns.

Enables broader real-world application: The weakly supervised approach makes the technology applicable to a "wider range of real applications," suggesting that responsible AI practices that reduce labeling demands can help scale beneficial AI systems to domains previously unable to afford full annotation.