Journal Article

1129 Embedding-guided patch selection improves early model confidence in multiplex tissue imaging

Weirather, Jason L, Singleterry Will, Jiang Sizun, Kaur Gulpreet

2025

Overview

BackgroundRobust and accurate classification of protein expression on individual cells is a critical step in automated and scalable analysis of multiplex tissue images. Achieving expert-level reliability requires well-trained computational models, but accurate cell type annotation remains a major bottleneck. Optimizing which training patches are labeled offers a practical and efficient route to resolve this. Here, we evaluated whether enforcing visual diversity during patch selection improves model calibration when labeled data are scarce.MethodsTwo sampling strategies were compared on human tonsil sections imaged with the Akoya PhenoCycler-Fusion platform, using CD11c as a reference marker:1) R-signal - signal-weighted random sampling based solely on CD11c intensity, and2) ER-signal - the same intensity weighting applied within 40 clusters generated from DINOv2 embeddings.Patch sets of 500-3500 examples were drawn from two independently imaged sections, and a convolutional neural network was trained five times per condition. Performance on held-out slides was measured with balanced accuracy, macro-F1, and the low-confidence rate (softmax < 0.60).ResultsER-signal lowered low-confidence predictions when data were limited. At 500 training patches, the low-confidence rate was 14% for ER-signal versus 26% for R-signal. At 1000 patches, it was 8% versus 10%. Above 2000 patches the rates converged to ≈ 5%.Overall accuracy was comparable. Both strategies reached similar balanced accuracy (~0.81) and macro-F1 (~0.78) by 2500-3000 patches; differences were minor and variable.ConclusionsEmbedding-guided patch selection improves early model confidence without compromising final accuracy, cutting uncertain calls by roughly half when only a few hundred labels are available. As training sets grow, the benefit diminishes, but the early gains underscore the value of embedding-aware sampling for annotation-efficient, high-plex image analysis. These results highlight the benefits of embedding-guided patch selection in multiplexed spatial proteomics images.

• Publisher: BMJ Publishing Group LTD
• Subject: Accuracy