Computational Pathology

The workhorse architecture for spatial feature extraction in pathology images. CNNs apply hierarchical filters to detect edges, textures, and complex morphological patterns.

• ResNet-50 and EfficientNet backbones commonly used for patch-level classification
• Transfer learning from ImageNet pre-training accelerates convergence on histology tasks
• Outputs feature vectors that feed into downstream MIL aggregators or segmentation heads
• Limitations include a fixed receptive field, making it difficult to capture long-range tissue architecture dependencies

Applies the self-attention mechanism to sequences of non-overlapping image patches, enabling the model to learn global spatial relationships across entire tissue regions.

• Divides a histology patch into 16×16 pixel tokens and processes them as a sequence
• Self-attention weights capture long-range dependencies between distant tissue structures
• Foundation models like UNI and Virchow use ViT architectures pre-trained on millions of WSIs
• Attention heatmaps provide built-in interpretability by highlighting diagnostically relevant regions

A weakly-supervised learning paradigm that trains models using only slide-level labels without requiring expensive pixel-level annotations for gigapixel whole-slide images.

• Treats each WSI as a bag of patches and the slide label as the bag label
• Attention-based MIL learns to weight diagnostically relevant patches while ignoring irrelevant tissue
• Enables training on large cohorts where only patient outcomes or diagnoses are available
• CLAM and TransMIL are widely adopted open-source MIL frameworks for computational pathology

An encoder-decoder architecture with skip connections that produces pixel-level tissue maps, essential for quantifying tumor regions, stroma, and cellular structures.

• The contracting path captures context; the expanding path enables precise localization
• Skip connections concatenate encoder features with decoder layers to preserve fine spatial detail
• Hover-Net extends this concept by predicting gradient maps to separate touching nuclei instances
• Foundation for tasks like Gleason pattern segmentation and TIL density quantification

Models tissue architecture as a graph where cells are nodes and spatial proximity defines edges, capturing the tumor microenvironment topology beyond pixel-based approaches.

• Each node encodes nuclear morphology, biomarker expression, or transcriptomic features
• Message passing between neighboring cells aggregates local microenvironment context
• Excels at modeling cell-to-cell interactions and spatial community detection in multiplex immunofluorescence data
• Enables discovery of prognostic spatial signatures invisible to human pathologists

Pre-trains models on massive unlabeled histology datasets by solving pretext tasks, learning generalizable visual representations without manual annotation.

• Contrastive learning (SimCLR, MoCo) pulls augmented views of the same patch together in embedding space
• Masked autoencoders (MAE) reconstruct intentionally hidden image patches, forcing the model to learn tissue semantics
• Foundation models pre-trained via SSL on millions of WSIs achieve state-of-the-art performance with minimal fine-tuning
• Reduces annotation burden by 10-100× compared to fully supervised approaches

A weakly-supervised learning paradigm essential for gigapixel pathology images where only a slide-level label (e.g., 'cancer present') is available, but no pixel-level annotations exist. The WSI is decomposed into thousands of patches (instances) packed into a bag. The model learns to aggregate patch-level predictions to make a slide-level classification, often using attention-based pooling to identify which regions contributed most to the decision.

A neural architecture that applies the self-attention mechanism to sequences of flattened image patches, capturing long-range spatial dependencies crucial for understanding tissue architecture. Unlike CNNs with their local receptive fields, ViTs model relationships between distant regions—such as tumor budding at the invasive front and the stromal reaction—in a single global operation. Foundation models like UNI and Virchow are ViT-based.

A critical pre-processing step that standardizes the color appearance of histology images to reduce domain shift caused by:

• Different staining protocols and reagent vendors
• Scanner-to-scanner color response variation
• Tissue preparation and fixation differences

Techniques range from histogram matching to generative adversarial networks (GANs) that learn a stain-independent representation, ensuring a model trained at one site generalizes to another.

The high-throughput extraction and mining of hundreds of quantitative features from digital pathology images to characterize tumor heterogeneity. Features include:

• Morphological: Nuclear size, shape, and texture
• Spatial: Cell clustering, nearest-neighbor distances
• Textural: Haralick features, Gabor filter responses

Pathomics transforms subjective visual assessment into objective, reproducible numerical descriptors for machine learning.

A large-scale pre-trained model, such as UNI, Virchow, or Prov-GigaPath, trained on massive histology datasets using self-supervised learning. These models learn general-purpose visual features from millions of unlabeled WSIs without requiring manual annotation. The resulting encoder can be fine-tuned with limited labeled data for diverse downstream tasks—from cancer subtyping to biomarker prediction—dramatically reducing the annotation burden.