UniCell: Universal Cell Nucleus Classification via Prompt Learning

Junjia Huang, Haofeng Li, Xiang Wan, Guanbin Li

Introduction

In computational pathology, the identification and classification of cell nuclei are foundational for cancer diagnosis and grading. However, a significant hurdle persists: annotation inconsistency, often referred to as the "label mess." Different datasets frequently use varying terminologies to describe identical biological entities. For instance, one dataset may label a cell as "Epithelial," while another classifies it as "Tumor." Furthermore, some datasets utilize broad categories like "Stromal," whereas others provide granular labels such as "Connective" or "Fibroblast."

The challenge of inconsistent labels across different pathology datasets.

This lack of standardization has historically forced researchers to train models on single datasets, preventing AI from learning "common knowledge"—the shared biological features that remain constant regardless of human naming conventions. UniCell addresses this by providing a universal framework capable of learning from multiple, inconsistently annotated datasets simultaneously.

Methodology

UniCell is an end-to-end architecture built upon the DEtection TRansformer (DETR) framework. Unlike traditional methods that rely on complex post-processing, UniCell directly predicts nucleus centroids and types. The core innovation of the framework is the Dynamic Prompt Module (DPM).

The UniCell architecture, featuring the Dynamic Prompt Module and dataset-specific prediction heads.

The DPM operates through several key mechanisms:

  1. Textual Prompts: Dataset names and specific cell categories are converted into learnable mathematical embeddings.
  2. Category Memory Bank: The model maintains a centralized repository of knowledge for all possible cell types across all integrated datasets.
  3. Localized Attention: To manage the fact that any given dataset only contains a subset of all possible cell types, the DPM employs an attention mask. This ensures the model focuses exclusively on relevant category information for the specific image being processed.
  4. Feature Enhancement: These prompts refine image features at the pixel level, assisting the model in distinguishing fine-grained details across diverse tissue types.

To accommodate different label sets, UniCell utilizes shared internal layers but employs separate "prediction heads" for each dataset, allowing the model to output the specific "native language" of each benchmark during training and inference.

Experimental Setup

The researchers evaluated UniCell across four diverse and challenging benchmarks:

  • CoNSeP: Colorectal adenocarcinoma images with three merged classes (inflammatory, epithelial, stromal).
  • MoNuSAC: Multi-organ nuclei from 37 hospitals, categorized into four types.
  • Lizard: A massive dataset of colonic tissue containing nearly half a million labeled nuclei across six categories.
  • OCELOT: Focused on the spatial relationship between background cells and tumor cells.

Results and Discussion

UniCell achieved new state-of-the-art (SOTA) results on all four benchmarks. In detection tasks, it improved the F-score by up to 4.7% over previous leading methods. The gains in classification were even more pronounced, with average F-score improvements ranging from 4% to 12.2%.

Key insights from the study include:

  • The Power of Shared Knowledge: Concurrent training on multiple datasets proves superior to individual training or sequential fine-tuning. The model develops a more robust representation of nuclear morphology by observing a wider variety of data.
  • Pixel-Level Enhancement: Empirical testing showed that using prompts to enhance image features (pixel-level) is significantly more effective than enhancing object queries. This suggests that prompts improve the model's fundamental perception rather than just its final decision-making.
  • Interference Mitigation: The localized attention mechanism effectively prevents "confusion" from non-existent categories in specific datasets, mitigating cross-dataset interference.

Conclusion

UniCell represents a major advancement in creating universal models for digital pathology. While the current framework focuses on centroid detection, it lays the groundwork for future "foundation models" in the field. Future work may involve extending the architecture to full pixel-wise segmentation and scaling the training to include vast amounts of unlabeled medical images through unsupervised learning.

Summary updated

BibTeX
@inproceedings{huang2024unicell,
  author = {Junjia Huang and Haofeng Li and Xiang Wan and Guanbin Li},
  title = {{UniCell: Universal Cell Nucleus Classification via Prompt Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i3.28009},
}
Figure