Multiple Instance Learning (MIL)

In MIL, training labels exist only at the bag level (the entire slide), not for individual instances (patches). The model must infer which patches are diagnostically relevant without explicit patch-level annotations.

• A positive bag contains at least one positive instance
• A negative bag contains only negative instances
• Eliminates the bottleneck of manual pixel-level annotation
• Enables training on routine clinical reports rather than curated datasets

MIL aggregators must be permutation-invariant: the output must not depend on the order of instances. This reflects the biological reality that tissue topology is not sequential.

• Mean pooling: Averages all instance features
• Max pooling: Selects the single most activated instance
• Attention-based pooling: Learns a weighted sum where weights sum to 1
• Transformer aggregation: Uses self-attention without positional encoding

Modern MIL architectures use gated attention mechanisms to learn instance-level importance weights. The model assigns higher attention to diagnostically relevant regions while suppressing irrelevant tissue.

• Attention weights provide built-in interpretability
• Generates attention heatmaps for pathologist review
• Gated attention uses a sigmoid gate with tanh activation
• Enables discovery of novel morphological biomarkers

Before aggregation, each patch is encoded into a compact feature vector using a pre-trained encoder. This reduces gigapixel images to manageable representations.

• ResNet-50 or ViT backbones pre-trained on histology
• Features typically 512 to 2048 dimensions per patch
• Self-supervised pre-training (e.g., SimCLR, DINO) improves features
• Foundation models like UNI and Virchow provide state-of-the-art embeddings

MIL relies on specific assumptions about the relationship between instances and bag labels. Understanding these is critical for model design.

• Standard assumption: A bag is positive if at least one instance is positive
• Threshold-based: A bag is positive if a minimum fraction of instances are positive
• Count-based: Bag label depends on the number of positive instances
• In pathology, the presence of even a small tumor region makes the slide positive

Before MIL training, whole-slide images undergo a standardized pre-processing workflow to extract viable tissue patches.

• Tissue segmentation: Otsu thresholding or U-Net to identify tissue regions
• Tiling: Extract non-overlapping patches at a fixed magnification (e.g., 20x)
• Quality control: Discard patches with blur, pen marks, or low tissue content
• Stain normalization: Standardize H&E color distributions across slides

A learning paradigm where models are trained using coarse, imprecise labels rather than exhaustive per-instance annotations. In digital pathology, this means using only slide-level diagnoses (e.g., 'cancer present') to train a model that must learn which individual patches contain tumor morphology. This dramatically reduces the annotation burden on pathologists compared to pixel-level segmentation.

A permutation-invariant aggregation operator that computes a weighted average of instance embeddings, where weights are learned by a gated attention mechanism. Key properties:

• Weights sum to 1, providing a natural interpretability mechanism
• The attention network learns to assign high weights to diagnostically relevant patches
• Produces an attention heatmap that highlights tumor regions without pixel-level supervision

Two fundamental MIL paradigms:

• Instance-level: Classifies each patch independently, then aggregates predictions (e.g., max pooling). Prone to noisy gradients when most patches are negative.
• Embedding-level: Maps each patch to a feature vector, aggregates all vectors into a single bag representation, then classifies the bag. Modern attention-based methods use this approach for superior performance and richer representations.

A mathematical property requiring that the aggregation function produce the same output regardless of input ordering. Since patches in a WSI have no natural sequence, MIL pooling operators must satisfy: f(x₁, x₂, ..., xₙ) = f(x_π(1), x_π(2), ..., x_π(n)) for any permutation π. Attention-based weighted averaging and Deep Sets architectures are designed to guarantee this property.