Accurate Boundary Delineation from Image-Level Tumor Labels

Establish accurate tumor-boundary delineation from coarse and noisy class-activation-map pseudo-labels generated using image-level supervision for weakly supervised medical image segmentation under client-specific missing MRI modalities.

Background

The paper identifies weakly supervised segmentation as a way to reduce the cost of expert voxel-level annotation by using image-level labels indicating whether a tumor is present. However, class activation maps generated from such labels provide coarse and noisy spatial pseudo-labels rather than precise tumor masks. This limitation is especially consequential when MRI modalities are missing at particular federated clients, because absent discriminative channels can further degrade localization and boundary quality. The paper therefore treats precise boundary delineation as an unresolved challenge motivating its refinement stage.

References

Class activation maps (CAMs) and their variants recover spatial localization from such labels , but the resulting pseudo-labels are coarse and noisy, making accurate boundary delineation a central open challenge.

— MOSAIC: Modality-agnostic Spectral Alignment for Federated Image-level Weakly Supervised Tumor Segmentation under Client-specific Missing Modalities  (2608.19788 - Garg et al., 20 Aug 2026) in Section 1, Introduction