MoA: Mixture of Aggregators Improves Slide-Level Diagnosis in Computational Pathology

Abstract

Multiple instance learning (MIL) is the standard for learning slide-level representations from whole slide images (WSIs), typically using a single attention-based aggregator to pool instance features. However, a single aggregator can struggle to capture morphological and compositional patterns of cells in pathology and cytology data, and different diseases may demand different pooling behaviours. We propose a mixture-of-aggregators framework that models complementary aspects of instance distributions in histology and hematologic cytology. A router with top-2 gating dynamically selects the most relevant aggregators per slide, and their outputs are fused into a patient-level representation. To avoid collapse to a single dominant expert aggregator, we add a load-balancing loss and Gumbel noise on the router logits. We evaluate our method on 19 different tasks from 16 datasets including histology and hematologic cytology. Compared to single-aggregator baselines, our approach improves diagnostic prediction accuracy by an average of 4.5% over ABMIL and 12.6% over TransMIL across all tasks. Beyond performance, our analysis shows that different aggregators attend to distinct, disease-specific instance distributions, providing interpretable insights into the diagnostic process.