Ilmu Komputer & AI editorial
Open AccessOA2026
MorphoSHAP: Rethinking the Unit of Attribution in Explanation for Deep Visual Models
A model-agnostic post-hoc method that uses morphological shapes as Shapley players for structured visual explanations
Anirudh Prabhakaran; Alexandre Rocchi; Gianni Franchi· 2026· DOI 10.48550/arXiv.2609.25815
The core problem
Visual attribution methods for deep models typically operate on pixels, superpixels, or regular patches. While these units can localize important regions, they offer limited insight into the structure of the evidence—such as whether it is a boundary, a texture, or a specific object part. MorphoSHAP addresses this gap by rethinking the unit of attribution: instead of pixels or patches, it uses morphological shapes derived from the Tree of Shapes. Each shape is characterized by its scale, geometry, and signed contribution, enabling explanations that answer where the evidence lies, what type of structure carries it, and how strongly it affects the prediction. This shared morphological vocabulary supports spatial, textual, and global class-level explanations beyond image-specific heatmaps. To the best of our knowledge, MorphoSHAP is the first SHAP-based image attribution framework to combine these different forms of explanation.
Innovation
The authors evaluate MorphoSHAP on five diverse datasets and three architectures. They report strong insertion/deletion performance, a common metric for attribution quality where pixels or regions are progressively added or removed based on attribution scores and the change in model output is measured. MorphoSHAP outperforms competing attribution methods on several benchmarks. For insertion, higher area under the curve (AUC) indicates better attribution; for deletion, lower AUC is better. The paper states that MorphoSHAP achieves strong performance on both, though specific numerical results are not provided in the abstract. Additionally, a user study shows that MorphoSHAP provides explanations that are easy to use and are preferred over standard attribution baselines. The study likely involved participants performing tasks such as identifying the most important shape or comparing explanations, with MorphoSHAP receiving higher preference ratings.
Visual attribution methods for deep models typically operate on pixels, superpixels, or regular patches. While these units can localize important regions, they offer limited insight into the structure of the evidence—such as whether it is a boundary, a texture, or a specific object part. MorphoSHAP addresses this gap by rethinking the unit of attribution: instead of pixels or patches, it uses morphological shapes derived from the Tree of Shapes. Each shape is characterized by its scale, geometry, and signed contribution, enabling explanations that answer where the evidence lies, what type of structure carries it, and how strongly it affects the prediction. This shared morphological vocabulary supports spatial, textual, and global class-level explanations beyond image-specific heatmaps. To the best of our knowledge, MorphoSHAP is the first SHAP-based image attribution framework to combine these different forms of explanation.
MorphoSHAP is a model-agnostic post-hoc method that formulates attribution as a Shapley game where the players are morphological shapes extracted via the Tree of Shapes. The Tree of Shapes provides a hierarchical representation of an image, where each node corresponds to a shape (a connected component of a level set) and is nested according to inclusion. For a given image, the set of shapes
is used as the player set in a cooperative game. The prediction function is evaluated on masked images where shapes are either kept or removed. The Shapley value for shape is computed as:
Why it matters
MorphoSHAP's key innovation is the use of morphological shapes as the unit of attribution, which provides a more structured and interpretable representation than pixels or patches. By leveraging the Tree of Shapes, it captures hierarchical and geometric information that is often lost in other methods. The signed Shapley contributions allow for both positive and negative evidence, giving a more complete picture of the model's decision. The shared morphological vocabulary enables different forms of explanation: spatial (heatmaps of shape contributions), textual (descriptions of shape properties), and global (class-level prototypes). This combination is unique among SHAP-based image attribution frameworks. The user study confirms that these explanations are not only more informative but also more usable. However, the method's reliance on the Tree of Shapes may introduce computational overhead, and the Monte Carlo approximation of Shapley values could be noisy. Future work could explore more efficient sampling or alternative morphological representations. Overall, MorphoSHAP represents a significant step towards more structured and interpretable visual explanations.
Who should read this
CS practitioners and researchers
Opening member content…