Exploring Unsupervised Cell Recognition with Prior Self-activation Maps

Pingyi Chen, Chenglu Zhu, Zhongyi Shui, Jiatong Cai, Sunyi Zheng, Shichuan Zhang, Lin Yang

Abstract

Deep learning has significantly advanced medical imaging, yet its reliance on large-scale, manually annotated datasets remains a bottleneck, particularly in histopathology. This paper presents a novel, label-free framework for cell recognition that leverages Prior Self-activation Maps (PSM). By utilizing the internal gradients of self-supervised neural networks, the framework generates high-quality pseudo-masks that serve as training targets for downstream tasks. Experimental results on the MoNuSeg and BCData benchmarks demonstrate that this unsupervised approach achieves performance competitive with, and in some cases superior to, supervised and weakly-supervised models.

Introduction

In histopathology, the density and complexity of cellular structures present a significant challenge for automated analysis. A single image patch can contain hundreds of individual cells, making manual segmentation or even point-based detection labor-intensive, expensive, and susceptible to inter-observer variability. This "label hunger" has limited the scalability of deep learning models in clinical settings.

While unsupervised methods like traditional clustering exist, they often fail to capture the semantic nuances of biological tissue. Furthermore, many existing AI-based unsupervised methods rely on domain adaptation, which still requires labeled data from a source domain. This research introduces a truly label-free pipeline that trains cell recognition models from scratch using only raw, unlabeled images.

Methodology

The proposed framework is centered on the generation of Prior Self-activation Maps (PSM) and follows a multi-stage pipeline:

1. Self-Supervised Proxy Task

An activation network (Res2Net101) is first trained without labels. The network performs a "proxy task"—specifically, learning to recognize if two different augmentations of the same image are similar. This forces the network to learn meaningful feature representations of the underlying cellular structures.

2. Generating Prior Self-activation Maps (PSM)

Once the network is trained, the authors extract gradients from its shallow layers. Because cells possess semi-regular shapes and dense distributions, the network naturally activates on these local features. By aggregating these gradient-weighted features, the system creates a PSM—a heat map highlighting potential cell locations.

3. Semantic Clustering Module (SCM)

To refine these fuzzy heat maps into crisp "pseudo-masks," the PSM is combined with the original raw image data and processed through a K-Means clustering algorithm. This step effectively separates the image into foreground (cells) and background.

4. Downstream Training

These pseudo-masks act as the ground truth for training standard segmentation (ResNet-34) and detection networks. To handle the challenge of adjacent, touching cells, the authors utilize a Voronoi map technique, which helps the model learn to distinguish individual cell boundaries more effectively.

Key Insights and Architectural Decisions

The Importance of Shallow Layers

A critical finding of this research is that features from the shallowest layers of the network are the most effective for cell recognition. Shallow layers have a smaller receptive field, making them better suited for capturing the small, local details of individual nuclei. In contrast, deeper layers capture more abstract, global features that are less useful for precise cell localization.

Similarity vs. Contrastiveness

While contrastive learning (learning to distinguish between different images) is a dominant trend in self-supervised learning, the authors found that "similarity" tasks (learning that two views of the same image are identical) are more effective for medical imagery. Because the differences between classes in histopathology can be extremely subtle, contrastive learning can inadvertently hinder the model by forcing it to find differences where semantic similarities are more important.

Multi-Class Capability

Unlike many unsupervised methods that are limited to binary "cell vs. background" classification, this framework successfully tackles multi-class detection. By leveraging the semantic clustering of the PSMs, the model can distinguish between different cell types, such as positive and negative cells in tumor datasets.

Experimental Results

The framework was evaluated on two major histological benchmarks:

  • MoNuSeg (Segmentation): The unsupervised method achieved an Aggregated Jaccard Index (AJI) of 0.542. This outperformed a fully supervised Unet trained from scratch (0.511) and a weakly-supervised method using point annotations (0.496).
  • BCData (Detection): The model achieved an F1-score of 0.811. While slightly below the state-of-the-art fully supervised models (~0.86), it significantly outperformed other weak-supervision baselines in counting accuracy, as measured by Mean Average Error (MAE).

Conclusion and Future Work

The PSM framework demonstrates that high-quality cell recognition is possible without the need for manual annotations. By leveraging the inherent activation patterns of self-supervised networks, the method provides a scalable solution to the label hunger problem in pathology.

However, limitations remain, particularly regarding false positives caused by ambiguous cell boundaries. Future research will focus on more robust boundary-handling techniques and exploring the generalizability of the PSM framework to other medical imaging modalities, such as X-ray and MRI, where semi-regular biological structures are prevalent.

Summary updated

BibTeX
@inproceedings{chen2023exploring,
  author = {Pingyi Chen and Chenglu Zhu and Zhongyi Shui and Jiatong Cai and Sunyi Zheng and Shichuan Zhang and Lin Yang},
  title = {{Exploring Unsupervised Cell Recognition with Prior Self-activation Maps}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=nieGC_y6cOL},
}
Figure