A real-time GPU-accelerated parallelized image processor for large-scale multiplexed fluorescence microscopy data

Abstract

Highly multiplexed, single-cell imaging has revolutionized our understanding of spatial cellular interactions associated with health and disease. With ever-increasing numbers of antigens, region sizes, and sample sizes, multiplexed fluorescence imaging experiments routinely produce terabytes of data. Fast and accurate processing of these large-scale, high-dimensional imaging data is essential to ensure reliable segmentation and identification of cell types and for characterization of cellular neighborhoods and inference of mechanistic insights. Here, we describe RAPID, a Real-time, GPU-Accelerated Parallelized Image processing software for large-scale multiplexed fluorescence microscopy Data. RAPID deconvolves large-scale, high-dimensional fluorescence imaging data, stitches and registers images with axial and lateral drift correction, and minimizes tissue autofluorescence such as that introduced by erythrocytes. Incorporation of an open source CUDA-driven, GPU-assisted deconvolution produced results similar to fee-based commercial software. RAPID reduces data processing time and artifacts and improves image contrast and signal-to-noise compared to our previous image processing pipeline, thus providing a useful tool for accurate and robust analysis of large-scale, multiplexed, fluorescence imaging data.

Multimodal and spatially resolved profiling identifies distinct patterns of T-cell infiltration in nodal B-cell lymphoma entities

“…Raw TIFF images were processed using the RAPID pipeline [79] in Matlab (version R2020a) with the following settings: nCyc=51, nReg= number of regions imaged (depending on TMA), nTil=49, nZ= number of z-planes imaged (depending on TMA), nCh= [1,4], nTilRow=7, nTilCol=7, overlapRatio=0.3, reg_range=1:nReg, cyc_range=1:nCyc, til_range=1:nTil, cpu_num=depending on computer system used, neg_flag=1, gpu_id=depending on number of GPUs available, cyc_bg=1. After deconvolution (two iterations), best focal plane selection, lateral drift compensation, stitching of individual images and background subtraction, processed images were concatenated to hyperstacks.…”

A single-cell multi-omic and spatial atlas of B cell lymphomas reveals differentiation drives intratumor heterogeneity

“…We processed raw TIFF images using the RAPID pipeline 80 in Matlab with the default settings. Post-processing, images were concatenated to hyperstacks.…”

Highly multiplexed spatial profiling with CODEX: bioinformatic analysis and application in human disease

“…Over the past 5 years, various computational methods and software tools have been developed to facilitate each of these analytical steps. Most carry out only one of the steps (CODEX Uploader [ 5 ], RAPID [ 25 ], CellSeg [ 26 ], Mesmer [ 27 ], CellPose [ 28 ], CELESTA [ 29 ], Astir [ 30 ], CytoMAP [ 31 ], histoCAT [ 32 ], TissueSchematics [ 33 ], and MISTy [ 34 ]), but a few attempt to address multiple steps or the entire workflow (Cytokit [ 35 ], MCMICRO [ 36 ], SIMPLI [ 37 ]) (Table 2 ).…”

“…RAPID [ 25 ] was developed for accurate and fast analysis of large-scale CODEX imaging data by the Nolan lab. It reduces processing time by two to threefold compared to the CODEX Uploader.…”