Abstract
Deep learning models for nuclei segmentation in histopathology are crucial for cancer diagnosis but suffer from catastrophic forgetting when learning new cell types incrementally. This paper proposes a novel framework for Incremental Semantic Segmentation (ISS) that addresses this through two primary mechanisms: Future-class Awareness (FCA) and Compatibility-inspired Distillation (CID). FCA mines unlabeled nuclei from the background to reserve representation capacity, while CID ensures seamless knowledge transfer between model versions. Experimental results on MoNuSAC and CoNSeP benchmarks show that the proposed method significantly outperforms existing state-of-the-art approaches, particularly in challenging sequential learning scenarios.
Introduction
In clinical pathology, the ability to update segmentation models with new cell classes without access to original training data is essential due to privacy regulations and data storage costs. However, standard incremental learning faces "catastrophic forgetting," where new information overwrites old knowledge. Furthermore, histopathological images present a unique "background shift" challenge: cell types to be learned in the future are initially labeled as background, training the model to actively ignore them. This paper introduces a framework designed to mitigate these issues by identifying potential future classes early and ensuring architectural compatibility during the distillation process.
Methodology
The proposed framework tackles incremental learning through a multi-pronged approach involving feature mining and structured distillation.
Future-class Awareness Mechanism (FCA)
Instead of treating all non-labeled pixels as background, the FCA mechanism utilizes class centroids (prototypes) in the feature space. By calculating the similarity between background pixels and these known prototypes, the model identifies regions that visually resemble nuclei but lack current labels. These regions are assigned an "unknown" label, preventing the model from becoming overconfident in classifying them as empty space and effectively "pre-learning" features for future classes.
Compatibility-inspired Distillation (CID)
To prevent forgetting, knowledge distillation is employed to align the new model with the old one. CID improves upon standard distillation by expanding the old model’s prediction range to match the new model's expanded class set. This "compatible interface" resolves alignment conflicts where the old model might have predicted "background" for a region the new model now correctly identifies as a "new class."
Local Pooled Outputs Distillation (POD)
To maintain spatial consistency, the framework incorporates multi-scale local distillation. This ensures that internal feature representations do not drift significantly across different layers of the network during incremental updates.
Experimental Results
The framework was evaluated on the MoNuSAC and CoNSeP datasets, focusing on the difficult "1-1" setting where four classes are learned sequentially.
Quantitative Performance
On the MoNuSAC 1-1 task, the proposed method achieved a mean Dice (mDice) score of 69.44, surpassing previous state-of-the-art (SOTA) methods like PLOP and MiB. Crucially, the model maintained a high score of 68.11 on "Old" classes, demonstrating a significant reduction in catastrophic forgetting compared to methods like ILT, which saw scores drop to 33.46.
Qualitative Analysis
Visual results indicate that the model maintains clean boundaries and correctly identifies multiple cell types even after four stages of incremental training. While baseline models often begin to confuse different nuclei types or lose track of earlier classes, the proposed approach retains high precision across the entire class set.
Conclusion
This research presents a robust solution for incremental nuclei segmentation in histopathology. By mining future classes from the background and ensuring compatibility between model versions, the framework successfully balances stability and plasticity. While the method relies on visual similarities between cell types, it provides a powerful foundation for evolving diagnostic models in data-constrained clinical environments.
Summary updated