Aggregators

Multiple Instance Learning (MIL) aggregators for combining features from multiple instances.

Overview

The aggregators module provides various strategies for aggregating instance-level features into bag-level representations, commonly used in medical imaging tasks where multiple patches or regions need to be combined.

Available Aggregators

Simple Aggregators

• MeanPoolingAggregator - Averages features across instances
• MaxPoolingAggregator - Takes maximum values across instances

Attention-Based Aggregators

• AttentionAggregator - Uses attention weights to aggregate instances
• GatedAttentionAggregator - Gated attention mechanism for instance aggregation

Advanced Aggregators

• DeepSetsAggregator - Deep Sets architecture for permutation-invariant aggregation
• TransformerAggregator - Transformer-based aggregation with multi-head attention

Unified Interface

The MILAggregator class provides a unified interface to all aggregation strategies.

Usage Examples

Basic Usage

from med_core.aggregators.mil import MILAggregator

# Create aggregator
aggregator = MILAggregator(
    input_dim=512,
    strategy='attention',
    attention_dim=128
)

# Aggregate features
features = torch.randn(4, 10, 512)  # [batch, instances, features]
aggregated = aggregator(features)    # [batch, features]

With Attention Weights

# Get attention weights for interpretability
aggregated, attention = aggregator(features, return_attention=True)
print(attention.shape)  # [batch, instances, 1]

Different Strategies

# Mean pooling
mean_agg = MILAggregator(input_dim=512, strategy='mean')

# Max pooling
max_agg = MILAggregator(input_dim=512, strategy='max')

# Gated attention
gated_agg = MILAggregator(input_dim=512, strategy='gated')

# Deep Sets
deepsets_agg = MILAggregator(
    input_dim=512,
    strategy='deepsets',
    hidden_dim=256,
    output_dim=256
)

# Transformer
transformer_agg = MILAggregator(
    input_dim=512,
    strategy='transformer',
    num_heads=8,
    num_layers=2
)