Multiplexed imaging drops the wet-lab compensation step: SpillOT denoises and corrects spillover zero-shot, surfacing hidden immune infiltration in diabetes data
Synopsis
Channel spillover in highly multiplexed imaging is normally corrected by extra wet-lab measurements of a compensation matrix, and random noise removal needs a representative training set. SpillOT learns only from the input image, using optimal-transport ideas to find similar patches across channels and keep only the true signal, performing denoising and spillover correction together. On Imaging Mass Cytometry, t-CyCIF and Opal multiplexed immunofluorescence data it surpasses existing zero-shot restoration methods and reveals previously undetectable immune infiltration patterns in diabetes data.
Interpretation
Restoration of highly multiplexed images can now happen entirely on the computational side: SpillOT trains and infers separately on each input image, needing neither extra wet-lab measurement of channel interactions nor a representative training set. Earlier routes either depended on paired or representative training data, or, for spillover, required imaging individual metal isotope spots on beads to determine spillover percentages and correct them. In the IMC pancreas test dataset, the Dy162-CD8 channel showed islet-shaped bleed-through; SpillOT removed that bleed-through and background noise while retaining CD8+ cells, without an experimentally generated metal spillover matrix.
Spillover is hard to remove because it exists as a proportion of the true signal and looks like legitimate biological structure rather than random noise. SpillOT treats spillover as a patch-similarity problem, building a comparison patch with pixel domino tiling (PDT, a checkerboard-style rearrangement that maps each pixel to a neighbor without duplicates) and then judging which channel's patch is more similar to it. In 25 ovarian cancer Opal multiplexed immunofluorescence images, Ki-67 contamination of the neighboring CD8 channel was removed, erasing nuclear CD8 while retaining membrane CD8; spillover removal took 272 s for a single 1,348×1,008 frame on a CPU.
Restored single-cell data move closer to known biological proportions and expose a layer of difference downstream analysis could not see. In the original Opal data, cells high in Ki-67 were almost universally labeled CD3−CD8+, a combination found only on natural killer cells, which would imply that nearly all proliferating cells are natural killer cells. After correction, the median number of CD3+ T cells per presumed CD8+CD3− natural killer cell rose from 0.58 to 8.6 (Wilcoxon rank-sum test, Bonferroni-adjusted P = 0.00106); across 189 diabetes pancreas images, the denoised data additionally showed a significant drop in CD8+ T cell density between the 2-year and 11-year post-diagnosis groups (P < 0.05).
Perspective
The demonstrated scope covers Imaging Mass Cytometry, Tissue Cyclic Immunofluorescence and Opal multiplexed immunofluorescence data, including pancreas, a psoriatic skin lesion, ovarian cancer and a public diabetes pancreas dataset; the authors recommend tiling large images and processing tiles in a distributed way on a cluster, and provide a Docker image and scripts runnable on Amazon Web Services.
SpillOT assumes uniform noise; if different image regions carry noise levels that a network cannot infer from the data, results would suffer, and the authors note that splitting the image into small patches could alleviate this but would draw on fewer sample points for training. Speed is another limit: the denoiser averaged 3,933.4 s per megapixel on an RTX A6000, versus 1.3 s per image for IMC-Denoise, though the latter needs 10,913 s of upfront training.
