AnnotateAnyCell: Open-Source AI Framework for Efficient Annotation in Digital Pathology

The paper introduces AnnotateAnyCell, an open-source semi-supervised framework that integrates active contrastive learning and human-in-the-loop feedback to reduce annotation time by 25% while achieving expert-level accuracy for cellular features in digital pathology.

The Gist

Imagine you are a master art critic trying to sort through a library containing millions of tiny, blurry photographs of cells. Your job is to look at each one and decide: "Is this cell dividing? Does it have a weird shape? Is the nucleus (the cell's brain) healthy or sick?"

Doing this manually for a whole slide of tissue is like trying to read every single book in a library to find a few specific sentences. It takes experts (pathologists) hundreds of hours, it's exhausting, and it's the biggest bottleneck stopping AI from helping doctors diagnose cancer faster.

Enter "AnnotateAnyCell."

Think of this new tool not as a robot that does the work for you, but as a super-smart, interactive librarian that helps you find the most important books to read first.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Needle in a Haystack"

In digital pathology, a single image (Whole Slide Image) is so huge it contains hundreds of thousands of cells. If a human has to look at every single one to train an AI, it's impossible.

• The Old Way: You read the books in order, from page 1 to page 1,000,000. It takes forever.
• The New Way (AnnotateAnyCell): You ask the librarian, "Show me the books that look most different from the ones I've already read."

2. The Magic Trick: The "Morphological Map" (UMAP)

The tool uses a special kind of math to create a 2D map of all the cells.

• Imagine you have a giant pile of mixed-up LEGO bricks.
• The tool sorts them on a table so that all the red bricks are in one pile, blue bricks in another, and weird-shaped bricks in a third.
• In the real world, this means cells that look similar (e.g., cells with "bubbly" nuclei) naturally group together on the screen.
• Why this helps: Instead of scanning the whole library, the pathologist just looks at one "pile" of similar cells and says, "Yes, these are all sick cells." The AI then learns that everything in that pile is likely sick too.

3. The "Guessing Game" (Active Learning)

This is the "Human-in-the-Loop" part.

1. The AI guesses: It looks at the unlabeled cells and makes a guess: "I think this cell is dividing."
2. The Human checks: The pathologist looks at the guess. If the AI is right, great! If the AI is wrong, the human corrects it.
3. The Smart Selection: The AI doesn't just ask for random corrections. It specifically asks for the cells it is unsure about or the ones that are rare.
   • Analogy: Imagine you are teaching a child to identify dogs. You don't show them 100 Golden Retrievers. You show them a Golden Retriever, then a Chihuahua, then a weird-looking mutt. You focus on the "hard" examples to teach them faster.

4. The Results: Speed and Accuracy

The researchers tested this on dog cancer samples (which are very similar to human bladder cancer).

• Time Saved: Using this tool, experts finished labeling 300 cells in 47 minutes. Doing it the old, sequential way took 63 minutes. That's a 25% time savings.
• Accuracy:
  • For spotting nucleoli (tiny structures inside the nucleus), the AI was 98% accurate.
  • For spotting mitotic figures (cells dividing), it was 96% accurate.
  • For cell shape, it was lower (around 60%), but that's because "shape" is very subjective even for humans.
• The "Small Sample" Miracle: The AI learned to recognize nucleoli with 95% accuracy after seeing only 215 examples. Usually, AI needs thousands of examples to learn this well.

5. Why This Matters

Currently, AI in medicine is stuck because we don't have enough "labeled" data (data where a human has already said, "This is cancer").

• AnnotateAnyCell is like a force multiplier. It allows a pathologist to do the work of three people in the same amount of time.
• It is Open Source, meaning any hospital or university can download it for free, unlike expensive commercial software.
• It builds trust. Because the human is always in the loop checking the AI's work, doctors are more likely to trust the final diagnosis.

The Bottom Line

AnnotateAnyCell is a smart assistant that organizes the chaos of millions of cells into neat, understandable groups. It asks the human expert only the questions they need to answer to learn the most, turning a months-long task into a few hours, and paving the way for AI to help diagnose cancer in hospitals that don't have huge budgets or armies of data scientists.

Technical Summary

1. Problem Statement

The integration of AI into digital pathology is hindered by the annotation bottleneck. Manual annotation of histopathological whole slide images (WSIs) requires hundreds of hours of expert time to delineate cellular boundaries and assign morphological classifications.

• Scale: Modern 40× WSIs contain hundreds of thousands of cells, often exceeding gigabytes in size.
• Limitations of Current Methods: Fully supervised learning requires massive labeled datasets. Existing semi-supervised or active learning tools often lack intuitive interfaces for pathologists, require heavy computational infrastructure, or operate at the "patch" level rather than the individual cell level, failing to leverage modern embedding spaces for interactive annotation.
• Generalization Issues: Models often fail to generalize across institutions due to variations in tissue preparation and scanner characteristics, necessitating costly retraining.

2. Methodology: The AnnotateAnyCell Framework

The authors propose AnnotateAnyCell, an open-source, semi-supervised framework that combines active learning, contrastive learning, and iterative human-in-the-loop feedback. The pipeline operates through four iterative stages:

A. Pre-processing and Segmentation

• Input: H&E-stained tissue images (40× magnification).
• Segmentation: Uses pre-trained Cellpose models (cyto3) to segment cells and nuclei.
• Multi-modal Representation: Each detected cell is extracted into three complementary representations:
  1. Raw image patches (128×128 for training, 512×512 for inspection).
  2. Isolated nuclear regions (surrounding tissue removed).
  3. Semantic masks.
• Dimensionality Reduction: UMAP is applied to 128×128 patches to create a 2D morphological feature space for visualization.

B. Interactive Active Learning Loop

• User Interface: Built with BokehJS, featuring a three-panel layout:
  • Central Panel: An interactive UMAP scatter plot where points represent cells, color-coded by status (unlabeled, labeled, selected).
  • Left Panel: Real-time analytics (histograms of nuclear area, N/C ratios).
  • Right Panel: High-resolution cell previews with annotation controls for specific morphological features.
• Workflow: Pathologists annotate representative cells in the UMAP space. The system uses uncertainty sampling with diversity promotion to select the most informative unlabeled samples for the next round of expert review.

C. Semi-Supervised Learning Engine

The core learning model integrates labeled and unlabeled data through:

1. Contrastive Learning: A convolutional backbone with a projection head is trained using a combination of supervised classification loss and InfoNCE loss. This embeds morphologically similar cells together while separating dissimilar ones.
   • Objective: $L_{total} = \alpha L_{CE} + \beta L_{InfoNCE}$
2. Pseudolabeling: The trained model generates class-balanced pseudolabels for unlabeled data by selecting the top-$N$ highest-confidence predictions per class to avoid bias toward dominant morphologies.
3. Multi-Modal Autoencoder: A convolutional autoencoder processes raw images, nuclear features, and masks simultaneously. It uses a Variational Autoencoder (VAE) framework with a variational bottleneck to learn compact latent embeddings (256-dimensional) that capture fine-grained morphology and context.
   • Objective: Combines reconstruction loss ($L_{recon}$) and feature prediction loss ($L_{feat}$).

D. Iterative Refinement

The loop iterates: Expert labels $\rightarrow$ Model retrains $\rightarrow$ Pseudolabels generated $\rightarrow$ UMAP visualization updates $\rightarrow$ Experts validate/correct. This progressively refines the morphological space and improves classification accuracy.

3. Dataset and Evaluation

• Dataset: Canine Invasive Urothelial Carcinoma (IncUC), a translational model for human bladder cancer.
• Scope: 5 Whole Slide Images (WSIs) at 40× magnification, covering low, intermediate, and high histological grades.
• Annotation Categories:
  • Mitotic figures (canonical/atypical).
  • Nucleoli characteristics (absent, single, multiple, prominent).
  • Chromatin texture (vesicular vs. hyperchromatic).
  • Nuclear shape (circular, oval, irregular).
• Annotators: 11 board-certified veterinary pathologists and residents.

4. Key Results

Efficiency Gains

• Time Reduction: Clustering-guided annotation required 47 minutes for 300 cells compared to 63 minutes for sequential annotation, representing a 25% reduction in time (95% CI 18–32%).
• Consistency: The time savings were consistent across different annotation volumes (22–30% reduction) without fatigue effects.

Classification Performance

Using 1,075 labeled samples, the framework achieved:

• Nucleoli Classification: 98.3% ± 1.4% accuracy. Notably, it reached 95.5% accuracy with only 215 samples.
• Mitotic Figures: 96.3% ± 1.2% accuracy.
• Nuclear Shape: 59.5% ± 2.1% accuracy (the most difficult category due to subjectivity).

Inter-Annotator Agreement

• High Agreement: Chromatin texture ($\kappa = 1.00$) and Nucleoli ($\kappa = 0.95$).
• Moderate/Low Agreement: Mitotic figures ($\kappa = 0.58$) and Nuclear shape ($\kappa = 0.36$), reflecting intrinsic morphological ambiguity in these categories.
• Feature Correlation: Weak pairwise correlations ($|r| < 0.3$) between features, supporting a multidimensional diagnostic model.

5. Key Contributions

1. Open-Source Framework: Provides a fully open-source, web-based tool for semi-supervised cellular annotation, democratizing access to AI-assisted pathology.
2. Novel Pipeline: Successfully integrates contrastive learning with UMAP-based interactive visualization, allowing pathologists to annotate based on learned feature representations rather than just spatial coordinates.
3. Data Efficiency: Demonstrates that high-accuracy models (e.g., for nucleoli) can be built with very few labeled samples (215) by leveraging unlabeled data and pseudolabeling.
4. Human-in-the-Loop Validation: Validates that clustering-based active learning significantly reduces cognitive load and annotation time compared to traditional sequential methods.

6. Significance and Future Impact

• Scalability: The framework offers a scalable path toward AI-assisted diagnostics, particularly in resource-constrained settings where expert time is limited.
• Clinical Utility: High-confidence predictions for well-defined features (nucleoli, mitotic figures) can be automated for screening, while ambiguous features (shape) can be flagged for expert review.
• Institutional Adaptation: The data efficiency allows institutions to rapidly adapt models to their specific tissue preparation protocols and scanners by annotating only 200–500 representative local cells.
• Future Work: The authors plan to address label noise tolerance for ambiguous morphologies, extend the framework to multi-tissue generalization, and validate clinical utility in prospective workflows.

In summary, AnnotateAnyCell bridges the gap between deep learning capabilities and clinical workflow realities, proving that combining expert knowledge with semi-supervised learning can achieve expert-level accuracy while substantially reducing the annotation burden.