MMNavAgent: Multi-Magnification WSI Navigation Agent for Clinically Consistent Whole-Slide Analysis

TL;DR

A clinically consistent Multi-Magnification WSI Navigation Agent (MMNavAgent) that explicitly models multi magnification interaction and adaptive magnification selection is proposed and a Cross-Magnification navigation Tool (CMT) that aggregates contextual information from adjacent magnifications to enhance discriminative representations along the navigation path is introduced.

Abstract

Recent AI navigation approaches aim to improve Whole-Slide Image (WSI) diagnosis by modeling spatial exploration and selecting diagnostically relevant regions, yet most operate at a single fixed magnification or rely on predefined magnification traversal. In clinical practice, pathologists examine slides across multiple magnifications and selectively inspect only necessary scales, dynamically integrating global and cellular evidence in a sequential manner. This mismatch prevents existing methods from modeling cross-magnification interactions and adaptive magnification selection inherent to real diagnostic workflows. To these, we propose a clinically consistent Multi-Magnification WSI Navigation Agent (MMNavAgent) that explicitly models multi magnification interaction and adaptive magnification selection. Specifically, we introduce a Cross-Magnification navigation Tool (CMT) that aggregates contextual information from adjacent magnifications to enhance discriminative representations along the navigation path. We further introduce a Magnification Selection Tool (MST) that leverages memory-driven reasoning within the agent framework to enable interactive and adaptive magnification selection, mimicking the sequential decision process of pathologists. Extensive experiments on a public dataset demonstrate improved diagnostic performance, with 1.45% gain of AUC and 2.93% gain of BACC over a non-agent baseline. Code will be public upon acceptance.

Figures (5)

• Figure 1: Top: Prior navigation methods typically use a fixed magnification, yielding isolated single-scale cues. Bottom: MMNavAgent explicitly models multi-magnification interaction and adaptive magnification selection via a closed MST-CMT loop: MST selects the next action, CMT extracts high-attention regions, and the loop repeats until Stop outputs the final multi-magnification regions.
• Figure 1: Comparison study of MMNavAgent for slide classification.
• Figure 2: MMNavAgent starts from thumbnail (step 0). At each step $t$, MST uses the accumulated memory to select the next action and invokes CMT to extract important regions at the chosen magnification, which are stored in memory. The loop ends (Final step) when MST selects Stop, outputting navigation results.
• Figure 3: Qualitative comparison of our predictions across five magnifications with eye-tracking annotations and two pathologists’ region annotations(EU and Asia).
• Figure 4: Pathology consistency and tumor overlap performance (%) for our agent