Feature Embedding

Compresses high-dimensional pixel data into a compact vector space (typically 128–1024 dimensions). This transformation preserves diagnostically relevant information while discarding noise, enabling efficient storage and computation.

• A 256×256 RGB patch (196,608 values) becomes a 512-dimensional vector
• Reduces memory footprint by 99.7%
• Enables real-time similarity search across millions of patches

Embeddings map morphologically similar tissue patterns to neighboring points in vector space. Tumor regions cluster together, while stroma, necrosis, and normal tissue form distinct neighborhoods.

• Cosine similarity measures morphological relatedness
• Enables unsupervised discovery of histological subtypes
• Foundation models like UNI and CTransPath produce highly separable embeddings

Embeddings learned via self-supervised pre-training on massive, unlabeled histology datasets serve as general-purpose feature extractors. These frozen representations transfer to downstream tasks with minimal labeled data.

• Pre-trained on millions of patches from diverse organs
• Fine-tuned for cancer grading, subtyping, and biomarker prediction
• Dramatically reduces annotation burden for rare disease applications

Positional information can be incorporated into embeddings to preserve tissue architecture context. Graph neural networks and vision transformers explicitly model spatial relationships between neighboring patches.

• Patch coordinates appended to feature vectors
• Enables modeling of tumor-stroma interfaces
• Critical for capturing architectural grading patterns like Gleason scoring

Embeddings are designed as interchangeable tokens for slide-level prediction. Multiple Instance Learning aggregators pool thousands of patch embeddings into a single whole-slide representation.

• Attention-based pooling weights diagnostically relevant regions
• CLAM framework uses clustering-constrained attention
• Enables weakly supervised learning from slide-level labels only

Pathology embeddings can be aligned with genomic and molecular representations in a shared latent space. This enables prediction of molecular biomarkers directly from routine H&E histology.

• Predicts microsatellite instability and tumor mutational burden
• Correlates morphological patterns with gene expression signatures
• Bridges the phenotype-genotype gap without costly sequencing

A weakly supervised learning paradigm where a model is trained on labeled bags of instances rather than individually labeled examples. In pathology, a WSI is treated as a bag containing thousands of unlabeled patches. The model learns to aggregate patch-level feature embeddings into a single slide-level prediction, enabling training from only slide-level diagnostic labels without requiring costly pixel-level annotations.

A pre-training strategy that learns visual representations from unlabeled histology images by solving pretext tasks. Common approaches include:

• Contrastive learning: pulling embeddings of augmented views of the same patch together while pushing others apart
• Masked image modeling: reconstructing intentionally hidden patch regions This produces rich, general-purpose feature embeddings that transfer effectively to downstream tasks with limited labeled data.

A neural network component that dynamically weights the importance of different input features. In MIL-based pathology models, attention operates over a set of patch embeddings, assigning higher weights to diagnostically relevant regions while suppressing irrelevant tissue. This produces an interpretable heatmap showing which tissue areas drove the classification decision, critical for clinical trust and auditability.

A transformer-based architecture that applies self-attention to sequences of image patches. Unlike CNNs that process local neighborhoods, ViTs model global relationships between all patch embeddings simultaneously. In computational pathology, ViTs capture long-range tissue architecture patterns that are diagnostically significant, achieving state-of-the-art performance on tasks ranging from cancer subtyping to biomarker prediction.

A deep learning architecture that models relationships between tissue patches as a graph structure. Each patch embedding becomes a node, and edges represent spatial adjacency or feature similarity. GNNs explicitly capture the spatial architecture of the tumor microenvironment, learning how the arrangement and interaction of different cell populations and tissue regions contribute to diagnosis and prognosis.