Lab Newsletter — September 15, 2026: The Art of Alignment

This week the digest lived in molecules — designing mRNA, inferring the conversations between cells, folding the genome. Today we come back to the lab’s imaging home, and to a step so routine it’s almost invisible: alignment. Before you can subtract one image from another, follow a cell across a movie, overlay forty rounds of multiplexed staining, or stack a hundred tissue sections into a three-dimensional whole, you first have to make the images line up. That is image registration — finding the spatial transformation that warps one image onto another — and for decades it meant running a fresh optimization for every single pair. Today’s digest is about how deep learning turned that slow, per-pair grind into a fast, learned function.

🧭 The founding move: registration as a function you learn once

Classical registration treats each image pair as its own optimization problem — accurate, but slow, and slower still for rich deformations. Balakrishnan et al. (IEEE TMI, 2019) changed the framing. Noting that “traditional registration methods optimize an objective function for each pair of images, which can be time-consuming for large datasets or rich deformation models,” they built VoxelMorph, which instead “formulate[s] registration as a function that maps an input image pair to a deformation field that aligns these images,” a function “parameterize[d] … via a convolutional neural network (CNN)” and trained once “on a set of images.” The payoff is amortization: “given a new pair of scans, VoxelMorph rapidly computes a deformation field by directly evaluating the function” — no per-pair optimization at all. And it costs nothing in accuracy, with “the unsupervised model’s accuracy … comparable to state-of-the-art methods, while operating orders of magnitude faster.” Registration became inference.

🧩 Where does the training signal come from? Image similarity itself

A learned function needs a training target — but you can’t hand-draw the correct deformation field for a pair of images. de Vos et al. (Medical Image Analysis, 2019) named the problem — “training of ConvNets for registration was supervised using predefined example registrations. However, obtaining example registrations is not trivial” — and answered it with the DLIR framework “for unsupervised affine and deformable image registration.” The trick is to borrow the objective from classical methods: DLIR’s networks “are trained … by exploiting image similarity analogous to conventional intensity-based image registration.” No labels required — the images supervise themselves. And by “stacking multiple of these ConvNets into a larger architecture,” DLIR does “coarse-to-fine image registration,” first the broad affine alignment, then the fine deformable warp.

🪢 Don’t tear the tissue: cycle-consistency for topology

A deformation field is powerful enough to fold anatomy onto itself or rip it apart — physically impossible warps that still lower the loss. Kim et al. (Medical Image Analysis, 2021) targeted exactly that failure: “the existing deep learning methods still have limitations in the preservation of original topology during the deformation with registration vector fields.” Their fix, CycleMorph, is elegant — if you warp image A onto B and then B back onto A, you should return to where you started. That round-trip constraint, “cycle consistency,” acts as “an implicit regularization to preserve topology during the deformation.” It keeps the warp physically sensible, and it’s practical at scale: CycleMorph “can be applied for both 2D and 3D registration problems” and “easily extended to multi-scale implementation to deal with the memory issues in large volume registration.” A geometric prior, enforced by construction.

🔭 Transformers, diffeomorphisms, and knowing when you’re unsure

Alignment is fundamentally about long-range correspondence — a landmark on one side of the image must find its partner across the frame — and that’s precisely where a CNN’s local receptive field strains. Chen et al. (Medical Image Analysis, 2022) made the case: “the performances of ConvNets may be limited by a lack of explicit consideration of the long-range spatial relationships in an image,” whereas transformers’ “substantially larger receptive field enables a more precise comprehension of the spatial correspondence between moving and fixed images.” Their TransMorph is “a hybrid Transformer-ConvNet model for volumetric medical image registration,” and it ships with two variants the lab especially values: “the diffeomorphic variants ensure the topology-preserving deformations, and the Bayesian variant produces a well-calibrated registration uncertainty estimate.” Not just a better alignment — a guaranteed-smooth one, and one that can tell you how much to trust it.

🎭 Registration without ever seeing a real image

Learned registration has a stubborn weakness: a model trained on one imaging contrast tends to fail on another. Hoffmann et al. (IEEE TMI, 2022) removed the dependency entirely. Observing that “learning-based techniques are fast at test time but limited to registering images with contrasts and geometric content similar to those seen during training,” SynthMorph learns “without acquired imaging data,” using “a generative strategy for diverse synthetic label maps and images that exposes networks to a wide range of variability, forcing them to learn more invariant features.” The result is a single contrast-agnostic model — and, remarkably, “training on arbitrary shapes synthesized from noise distributions results in competitive performance, removing the dependency on acquired data of any kind.” A registration network taught by pure synthesis, generalizing to modalities it never met — a striking cousin of the lab’s interest in simulation-driven learning.

🗺️ Aligning space itself: from slices to a 3D atlas

The same problem reappears the moment imaging meets omics. Zeira et al. (Nature Methods, 2022) tackled it for spatial transcriptomics, where each slice “measures mRNA expression across thousands of spots … while recording the two-dimensional (2D) coordinates of each spot.” Their PASTE — probabilistic alignment of ST experiments — “align[s] and integrate[s] ST data from multiple adjacent tissue slices,” computing “pairwise alignments of slices using an optimal transport formulation that models both transcriptional similarity and physical distances between spots.” From those pairwise maps it “construct[s] a stacked 3D alignment of a tissue,” and the integration pays off downstream: “the PASTE integrated slice improves the identification of cell types and differentially expressed genes.” Note the engine — optimal transport, the same primitive behind the spatial cell-cell-communication methods from two days ago. Alignment is what turns a pile of 2D slices into a 3D map.

🧫 Why it’s our kind of problem

Read across the six and the arc is quietly the lab’s own: reframe a slow optimization as a learned function (VoxelMorph), supervise it with the data’s own structure (DLIR), constrain it with a physically-meaningful prior (CycleMorph’s topology), upgrade the architecture for long-range correspondence and report uncertainty (TransMorph), erase the data dependency with synthesis (SynthMorph), and carry it into a new modality (PASTE). Registration is the hidden prerequisite of almost everything the lab images: the forty rounds of multiplexed staining have to align cycle-to-cycle; live-cell movies on the REEF imaging farm and Agent-Lens need drift correction and multi-position stitching; and an autonomous microscope that knows its own registration uncertainty (TransMorph’s Bayesian variant) can decide, on its own, when a frame needs re-imaging. These are open, benchmarked tools — VoxelMorph, TransMorph, SynthMorph and PASTE all ship public code, the same publish-the-model spirit behind the BioImage Model Zoo and BioEngine. And it’s a concrete rung toward the virtual cell: building a cell model from imaging means fusing many partial views — timepoints, channels, sections, modalities — into one coherent, registered whole. Alignment is the connective tissue of that data engine. Before you can model the cell, you have to line up everything you saw of it.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits; a hypha-search surrogate sweep surfaced only virtual-cell horizon items, nothing breaking. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.