MSCF-Net: unified multi-scale feature disentanglement for co-registered nucleus segmentation and virtual staining

Integrated single-cell morpho-molecular analysis is vital for precision pathology, yet accurate nuclear instance segmentation from standard histopathology images remains challenging due to staining variability and complex cellular morphology. While virtual staining provides complementary molecular cues, existing frameworks often treat generation and segmentation as independent or loosely coupled tasks, limiting the effective use of auxiliary information and leading to feature entanglement.

To address this limitation, we propose MSCF-Net, a unified multi-modal framework in which nuclear instance segmentation is the primary objective, and adversarially learned virtual staining serves as an auxiliary modality to guide and regularize the segmentation process. The model adopts a shared encoder with parallel branches and explicitly integrates auxiliary cues through three key designs: Multi-scale Differential Enhancement blocks that provide scale-consistent feature representations, an adversarially regularized task-aware gating mechanism that selectively emphasizes boundary-relevant features, and consistency-regularized auxiliary-guided skip connections that enable controlled integration of high-resolution spatial details.

Experiments on BCData and DeepLIIF, including comprehensive benchmarking against representative methods, demonstrate that MSCF-Net achieves strong and robust performance in nuclear instance segmentation while maintaining reliable virtual staining quality.

The results confirm that MSCF-Net effectively addresses the challenges of feature entanglement in multi-task learning. By leveraging adversarially learned virtual staining as a structured prior, the framework enhances the precision of instance segmentation in morphologically complex scenarios, demonstrating its potential as a robust tool for integrated single-cell analysis in digital pathology.