**Title: Achieving Single-Channel Tissue Segmentation Through Cross-Modal Distillation from Foundation Models

Abstract:

While multiplexed fluorescence microscopy enhances tissue segmentation by offering complementary data streams—such as nuclear (DAPI) and membrane (E-cadherin) markers—it creates richer spatial context than single-channel imaging. However, the necessity of all channels during inference restricts the deployment of multiplexed models in scenarios where only a limited subset of data is accessible. To address this, we introduce a cross-modal knowledge distillation framework. This approach transfers semantic insights from a frozen foundation model teacher, which processes multiplexed inputs, to a streamlined student model that functions using only the nuclear channel.

Our distillation strategy integrates MSE-based probability matching, boundary-aware supervision, and learnable uncertainty weighting. We assessed the efficacy of this method using SAM ViT-H and CellSAM as teachers, across four distinct U-Net student architectures: Swin-Tiny (27M parameters), ResNet18 (11M), EfficientNet-B0 (5.3M), and MobileNetV3 (1.5M). The evaluation was conducted on the TissueNet and BBBC038 datasets.

Results from TissueNet demonstrate that the Swin-Tiny student, distilled from SAM, achieved a Dice score of 78.36 (±1.44). This represents a significant 13.05-point improvement over the baseline without knowledge distillation (65.31 ±1.35) and recovers 87.9% of the teacher’s oracle performance (89.12 ±1.21), all while reducing the parameter count by 23 times. Knowledge distillation consistently boosted performance by approximately 12 Dice points across all four student models, underscoring the architecture-agnostic nature of the distillation process. Furthermore, SAM ViT-H proved superior to CellSAM as a teacher across all tested configurations. Additionally, cross-dataset testing on BBBC038 yielded consistent performance gains without the need for teacher retraining.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC