Slide-Level Classification

Processing a whole slide image requires a tessellation and feature extraction pipeline before classification can occur. A single WSI at 40x magnification contains billions of pixels, demanding efficient computational strategies.

• Patch extraction: The tissue region is divided into manageable tiles (typically 256x256 or 512x512 pixels)
• Tissue segmentation: Background glass and artifacts are filtered out to avoid processing empty regions
• Feature encoding: A pre-trained encoder (often a Vision Transformer or ResNet) converts each patch into a compact embedding vector
• Hierarchical processing: Some architectures process patches at multiple magnifications to capture both cellular detail and tissue architecture

Pathologists naturally examine tissue at multiple scales, and slide-level classifiers increasingly mimic this behavior through multi-scale architectures. Combining features from different magnifications captures both fine cellular morphology and broader architectural patterns.

• Low magnification (5x-10x): Captures tissue architecture, tumor-stroma relationships, and regional heterogeneity
• High magnification (20x-40x): Resolves nuclear atypia, mitotic figures, and cellular-level features
• Fusion strategies: Late fusion concatenates features from separate magnification-specific encoders; early fusion processes multi-scale inputs jointly
• Attention across scales: Cross-scale attention mechanisms allow the model to relate cellular features to their architectural context

Beyond the slide-level prediction, clinical deployment demands visual explanations of which tissue regions drove the classification. Attention scores and gradient-based methods provide spatial heatmaps that pathologists can review.

• Attention heatmaps: Directly visualize the learned patch importance weights as an overlay on the WSI
• Grad-CAM: Produces class-specific localization maps by backpropagating gradients through the feature encoder
• High-attention region review: Enables pathologists to rapidly navigate to the most diagnostically relevant areas
• Quality assurance: Mismatches between attention maps and known diagnostic criteria can flag model errors or out-of-distribution inputs

Slide-level classifiers must maintain accuracy across scanner types, staining protocols, and institutional variations. Domain shift—where training and deployment distributions differ—remains a critical barrier to clinical translation.

• Stain normalization: Preprocessing techniques like Macenko or Vahadane methods standardize color distributions across slides
• Scanner invariance: Models trained on one scanner vendor often degrade on others due to sensor and compression differences
• Multi-site training: Federated learning and domain adversarial training improve robustness to unseen institutions
• External validation: Regulatory clearance requires demonstrating performance on completely independent, multi-site cohorts

A neural network component that dynamically weights the importance of each patch in a WSI. Instead of averaging all patches equally, the model learns to focus on diagnostically relevant tissue.

• Gated attention: Introduces a sigmoid gate to learn non-linear feature interactions.
• Multi-head attention: Allows the model to attend to multiple morphological patterns simultaneously.
• Produces attention heatmaps that highlight tumor regions for pathologist review.

A compact numerical vector representing the morphological content of a pathology patch. Pre-trained foundation models or self-supervised encoders convert each 256x256 pixel tile into a 512- to 2048-dimensional vector.

• Reduces a gigapixel WSI to a sequence of ~10,000 feature vectors.
• Enables efficient slide-level aggregation without pixel-level processing.
• Modern pathology-specific encoders like UNI and Prov-GigaPath provide state-of-the-art representations.

The preprocessing step that tessellates a gigapixel WSI into manageable tiles for neural network ingestion. Tissue regions are identified via tissue segmentation, then extracted at a target magnification.

• Typical patch size: 256x256 pixels at 20x or 40x magnification.
• Background and artifact regions are discarded to reduce computational waste.
• Overlapping patches can be used to increase spatial coverage.
• Output: tens of thousands of patches per slide.

A large-scale pre-trained model that serves as a general-purpose feature extractor for downstream pathology tasks. Trained on massive, diverse histology datasets using self-supervised learning.

• UNI: A ViT-Large model trained on over 100 million patches from 100,000+ slides.
• Prov-GigaPath: A whole-slide foundation model pre-trained on 1.3 billion pathology patches.
• Eliminates the need for task-specific architecture design; simply extract embeddings and train a lightweight MIL aggregator.