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

Fatih Ibrahim Ozlugedik, Muhammed Furkan Dasdelen, Rao Muhammad Umer, Carsten Marr

Keywords: computational pathology, multiple instance learning

Abstract: Multiple instance learning (MIL) has become the standard approach for learning meaningful representations from whole slide images (WSIs) in computational pathology. A variety of aggregation strategies have been proposed to compress instance-level features into slide-level representations, with attention mechanisms often employed to capture diagnostically relevant instances. However, single aggregators struggle to capture the diverse morphological patterns present in heterogeneous pathology and cytology datasets, where different diseases may require distinct aggregation strategies to identify relevant instances. In this work, we introduce a framework that employs a mixture of aggregators to model complementary aspects of instance distributions in histology and cytology images. Through a router with top-2 gating, our architecture dynamically selects the most relevant aggregators for each slide, whose outputs are weighted and later fused into a patient-level representation for classification. To prevent collapse into a single dominant aggregator, we introduce a balancing loss and Gumbel noise on router logits that encourages effective utilization of multiple aggregators during training. We extensively evaluate our method on six diverse histology datasets from different organs and two white blood cell datasets. Compared to both attention-based MIL and transformer baselines, our approach improves diagnostic prediction accuracy by 3-4\% on average across all datasets. Beyond performance gains, our analysis shows that different aggregators attend to distinct, disease-specific instance distributions, providing interpretable insights into the diagnostic process. Thus, our method improves both the performance and interpretability of MIL models used in clinical pathology workflows, influencing diagnostic decision-making.

Primary Subject Area: Application: Histopathology

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 238