Miniature: Unsupervised glimpses into multiplexed tissue imaging datasets as thumbnails for data portals

Abstract

Multiplexed tissue imaging can illuminate complex spatial protein expression patterns in healthy and diseased specimens. Large-scale atlas programs such as Human Tumor Atlas Network and are relying heavily on highly-multiplexed approaches including CyCIF and CODEX to image up to 100 antigens. Such high dimensionality allows a deep understanding of cellular diversity and spatial structure, but can provide a challenge for image visualization and exploration. One challenge for data portals and visualization tools is the generation of an informative and pleasing image preview that captures the full heterogeneity of the image, rather than relying on a multi-channel overlay that may be restricted to 4-6 channels. We describe Miniature, a tool to automatically generate informative image thumbnails from multiplexed tissue images in an unsupervised and scalable manner. Miniature aims to aid researchers in understanding tissue heterogeneity and identifying potential pathological features without extensive manual intervention. Miniature uses a choice of unsupervised dimensionality reduction methods including uniform manifold embedding and projection (UMAP), t-distributed stochastic neighbor Embedding (t-SNE), and principal Component analysis (PCA) to reduce on-tissue pixels from a low-resolution, high dimensional image to two or three dimensions. Pixels are then color encoded by their coordinate in low dimensional space using a choice of color maps. We show that perceptually distinct regions in Miniature thumbnails reflect known pathological features seen in both the source multiplexed tissue image and H&E imaging of the same sample. We evaluate Miniature parameters for dimensionality reduction and pixel color encoding to recommend default configurations that maximize perceptual trustworthiness to both the low-dimensional embedding and high-dimensional image and provide high Mantel correlation between the perceived color difference (delta E 2000) and distance in high- and low-dimensional space. By simulating color vision deficiency, we show that Miniature thumbnails are accessible to all. We demonstrate that Miniature thumbnails are suitable for a wide range of multiplexed tissue imaging modalities and show their application in the Human Tumor Atlas Network Data Portal.