CARE: A Molecular-Guided Foundation Model with Adaptive Region Modeling for Whole Slide Image Analysis

The paper introduces CARE, a molecular-guided foundation model that utilizes a two-stage pretraining strategy to automatically partition whole slide images into biologically relevant, adaptive regions, achieving superior performance across diverse pathology tasks with significantly less pretraining data than existing models.

The Gist

The Big Problem: The "Pixelated" Pathologist

Imagine a pathologist looking at a massive, high-resolution photo of a tissue sample (called a Whole Slide Image or WSI). This image is so huge it's like looking at a city from a satellite.

Current AI models try to understand this image by chopping it up into tiny, perfectly square tiles (like a grid of pixels). They look at each square individually and then try to guess the diagnosis.

The Flaw: This is like trying to understand a novel by reading it one letter at a time, or understanding a city by looking at individual bricks.

• The Issue: Tissue isn't made of perfect squares. Tumors have weird, organic shapes. By forcing the image into a rigid grid, the AI cuts right through important structures, mixing up healthy cells with cancer cells. It's like trying to describe a cat by only looking at the square that contains its tail and the square that contains its ear separately. You lose the "big picture" of what the cat actually looks like.

The Solution: CARE (The Smart Organizer)

The researchers created a new AI model called CARE (Cross-modal Adaptive Region Encoder). Instead of forcing the image into a rigid grid, CARE acts like a smart, flexible puzzle solver.

Here is how it works, step-by-step:

1. The "Adaptive Region" (The Flexible Puzzle)

Instead of cutting the image into squares, CARE looks at the tissue and says, "Hey, this cluster of cells looks like a tumor, and that cluster looks like healthy tissue. Let's draw a custom shape around them."

• Analogy: Imagine you are organizing a messy room.
  • Old AI: Uses a grid of identical boxes. It forces a round lamp into a square box, crushing it.
  • CARE: Uses custom-shaped bags. It puts the lamp in a bag that fits its shape perfectly, and the books in a different bag. It respects the natural boundaries of the objects.

This allows CARE to group cells that actually belong together, making the AI's "understanding" much more accurate and easier for doctors to trust.

2. The "Molecular GPS" (The Secret Map)

This is the most unique part of CARE. Usually, AI learns just by looking at pictures. But CARE has a secret weapon: Molecular Data (RNA and protein profiles).

• The Analogy: Imagine you are trying to learn a new language just by looking at pictures of people. It's hard. But now, imagine you have a translator standing next to you who whispers the meaning of what you are seeing.
  • RNA/Proteins: These are the "whispers." They tell the AI, "That weird-looking cluster of cells? That's actually a specific type of cancer gene."
  • The Result: CARE uses this biological "GPS" to guide its shape-drawing. It learns to ignore irrelevant areas and focus intensely on the specific spots where the biology is happening. This is why it needs 10 times less data to learn than other models—it's not just guessing; it's being taught the "why" behind the "what."

3. The Two-Stage Training (The Internship)

CARE learns in two phases, like a medical student:

1. Stage 1 (The Visual Intern): It looks at thousands of tissue images without any labels, just learning to recognize shapes and textures on its own (Self-Supervised).
2. Stage 2 (The Expert Mentor): It gets paired with the "Molecular GPS" (RNA/Protein data). The mentor corrects the student, saying, "No, that shape isn't just a random blob; it's a specific biological region. Adjust your focus."

Why This Matters (The "So What?")

• Better Accuracy: In tests, CARE beat almost every other AI model on 33 different medical tasks, from spotting cancer types to predicting how long a patient might live.
• Data Efficiency: It achieved these top results using only 10% of the data other models require. This is huge because medical data is hard to get and expensive to label.
• Trustworthy: Because CARE draws "custom shapes" around the actual tissue structures (rather than random squares), doctors can look at the AI's heatmaps and actually understand why it made a diagnosis. It aligns with how human pathologists think.

Summary

CARE is a new AI for cancer detection that stops treating tissue like a grid of squares. Instead, it acts like a flexible artist, drawing custom shapes around the actual biological structures. It learns faster and better by using a "molecular GPS" (RNA and proteins) to guide its focus, resulting in a smarter, more reliable tool for doctors.

Technical Summary

1. Problem Statement

Computational pathology (CPath) has seen significant advances with foundation models trained on Whole Slide Images (WSIs). However, existing models face two critical limitations:

• Rigid Partitioning: Most models treat WSIs as collections of fixed-size, regular patches (similar to character-level tokens in NLP) or fixed grid regions. This approach ignores the heterogeneous, non-uniform organization of biological tissues, often splitting coherent tissue structures across patch boundaries.
• Lack of Biological Grounding: Current models rely heavily on visual features alone. They lack explicit mechanisms to align visual representations with underlying biological realities (e.g., gene expression or protein profiles), limiting their ability to identify biologically relevant Regions of Interest (ROIs) and reducing interpretability.
• Data Inefficiency: State-of-the-art models typically require massive pretraining datasets (hundreds of thousands of slides), which are expensive to curate and annotate.

2. Methodology: CARE Framework

The authors propose CARE (Cross-modal Adaptive Region Encoder), a foundation model designed to partition WSIs into irregular, morphologically coherent regions guided by molecular data.

A. Adaptive Region Modeling

Instead of fixed grids, CARE introduces an Adaptive Region Generator (ARG) that dynamically partitions patches into semantically coherent regions:

• Soft Inclusion & Assignment: Patches are assigned to regions based on a "soft inclusion" matrix derived from spatial proximity and feature similarity (cosine similarity between patch features and region descriptors).
• Top-K Selection: Each patch selects its top-3 candidate subregions. A weighted score combining semantic affinity and spatial proximity determines the final region assignment, creating irregular, shape-adaptive regions similar to "word-level tokens" in NLP.
• Region Structuring Loss (RSL): A novel loss function prevents model collapse (where all patches select the same region) by encouraging a diverse distribution of region assignments.

B. Two-Stage Pretraining Pipeline

CARE utilizes a data-efficient, two-stage pretraining strategy:

1. Stage I: Unimodal Self-Supervised Pretraining:
   • Uses the iBOT algorithm (a Vision Transformer-based teacher-student framework) on 34,277 WSIs without segmentation annotations.
   • Employs frozen feature augmentation (perturbing feature space rather than pixels) and clusters patches into sub-WSIs to handle gigapixel scales with larger batch sizes.
2. Stage II: Cross-Modal Contrastive Pretraining:
   • Aligns WSI embeddings with molecular profiles using a CLIP-style objective.
   • RNA Guidance: Aligns slide features with RNA expression profiles (using Hallmark gene sets) to provide broad, high-coverage supervision.
   • Protein Guidance: Further refines the model by aligning with protein expression profiles (top-10 abundant proteins) to incorporate high-specificity biological signals.
   • This molecular guidance sharpens region boundaries, ensuring the model focuses on biologically relevant tissue areas.

C. Slide-Level Aggregation (SPF)

To generate a slide-level embedding, CARE uses a Semantic and Prior Fusion (SPF) module:

• It aggregates adaptive region features using a weighted sum of Coverage Prior (proportion of patches in a region) and Semantic Attention (learned weights based on region features).
• The region with the highest aggregated weight is selected as the ROI, enabling both slide-level prediction and ROI-level analysis.

3. Key Contributions

1. Adaptive Region Generator: A novel architecture that replaces rigid patch grids with irregular, morphologically coherent regions, mimicking the way pathologists identify ROIs.
2. Molecular-Guided Pretraining: A curriculum learning strategy that leverages RNA and protein data to refine visual representations, significantly reducing the need for labeled data.
3. Data Efficiency: CARE achieves superior performance using only ~10% of the pretraining data typically required by mainstream foundation models (e.g., TITAN, PRISM).
4. Dual-Mode Output: The model supports both slide-level classification and explicit ROI selection, enhancing interpretability.

4. Experimental Results

The model was evaluated on 33 downstream benchmarks across 9 datasets, covering morphological classification, molecular prediction (gene mutation), and survival analysis.

• Performance: CARE achieved state-of-the-art (SOTA) or near-SOTA results across all categories.
  • Morphological Classification: Outperformed baselines (CHIEF, PRISM, GigaPath, etc.) in linear probing and fine-tuning settings.
  • Molecular Prediction: Showed a distinct advantage over other models, attributed to the molecular-guided pretraining.
  • Survival Analysis: Achieved the best C-index in multiple tasks, demonstrating strong prognostic capabilities.
• Efficiency: Despite using significantly less pretraining data (34k WSIs vs. 170k–580k for competitors), CARE maintained lower computational complexity (FLOPs) and parameter counts while outperforming larger models.
• Ablation Studies: Confirmed that removing the adaptive region mechanism or the molecular guidance stages significantly degraded performance.

5. Significance

• Paradigm Shift: CARE moves away from the "fixed patch" paradigm in computational pathology toward adaptive, biology-aware region modeling, aligning AI processing with human pathological workflows.
• Clinical Relevance: By explicitly identifying biologically relevant ROIs (validated against pathologist annotations), CARE improves model interpretability and trustworthiness for clinical deployment.
• Scalability: The data-efficient pretraining pipeline suggests that high-performance pathology foundation models can be built without the prohibitive cost of massive, fully annotated datasets, potentially democratizing access to advanced CPath tools.

In summary, CARE represents a significant step forward in computational pathology by integrating morphological adaptability with molecular guidance, achieving superior performance with greater efficiency and interpretability than existing foundation models.