Find real signal
Estimate a global signal-versus-background threshold for every marker channel.
CASE STUDY · SCIENTIFIC COMPUTING / GPU PIPELINE
BioSET
A modular, GPU-accelerated pipeline that transforms multi-channel 3D tissue imaging into a compact database of biological marker co-localization results.
01 · THE SCALE
Turn a tissue volume too large for memory into relationships researchers can explore.
A single imaging volume contains 70 biological marker channels across roughly 194 × 5,508 × 10,908 voxels. The raw data cannot be analyzed manually or loaded into memory as one dense object.
The pipeline processes 3,784 tiles independently, measures where markers occupy the same spatial regions, and feeds those results to the BioSET Visualizer for interactive heatmaps, overlap plots, and 3D exploration.
02 · THE PIPELINE
Six stages convert raw signal into spatial evidence.
Stages one through five run on GPUs; the final aggregation stage runs on CPUs. Each stage is an independent Nextflow process with clear inputs and outputs.
Simplify each tile
Convert intensity data into a clean map of where each marker is present.
Remove noise
Use connected-component filtering to discard small specks and preserve structures.
Measure nearness
Expand regions across five radii from 0 to 2 μm so close markers count meaningfully.
Score overlap
Measure shared geometry across marker pairs and small sets with overlap metrics.
Rebuild the volume
Combine per-tile measurements into a queryable result for the entire tissue volume.
03 · THE SYSTEM
Modularity makes expensive research workflows practical to iterate.
Every stage can be skipped, resumed, or used as a new starting point. A parameter change only reruns the affected downstream work rather than restarting the whole volume.
The final .bioset artifact is a gzip-compressed SQLite database containing overlap scores, dilation radii, channel metadata, and aggregate results—not the raw images themselves.
Large-scale imaging, reduced to evidence researchers can query.
BioSET runs end to end on NYU’s H200 GPU cluster, produces verified results on real tissue tiles, and emits a database the visualizer can load directly.