Hospitals hold decades of tumor specimens fixed in formalin and embedded in paraffin, each linked to the patient's clinical outcome. Methods such as Deep Visual Proteomics can now cut out a few hundred cells of one type from such a block and measure thousands of their proteins. To decide which cells to cut, however, they first stain the tissue so a microscope and a segmentation network can see the cells, and they have no measure of how much protein each cut region actually holds.
We propose replacing the stain with a label-free phase image. A deparaffinized, unstained tissue section is almost transparent. It barely absorbs light and mainly delays it, so each point acts like a pure phase shifter: its transmission is a complex number of modulus one. Radar and wireless engineers solve exactly this kind of problem, where every unknown must sit on the unit circle, and they have fast, analyzed algorithms for it. We will use those algorithms to reconstruct phase images of archival sections from a small number of cheap LED-array exposures. We will use the images in two ways: to segment cells for laser microdissection without a stain, and to estimate the dry protein mass of every region before it is cut, since the phase delay grows linearly with that mass.
Core Claim
The central claim is that, for unstained formalin-fixed sections, requiring every transmission value to have modulus one is a physically correct prior that roughly halves the number of unknowns. With that prior, we expect a usable phase image from far fewer illumination angles than an unconstrained complex reconstruction needs. This matters because imaging time competes directly with microdissection throughput, reported at about 1,250 contours per hour in Deep Visual Proteomics.
Once the image exists, it supports two downstream uses: segmentation for microdissection without a chemical stain, and a dry-mass estimate for each contour that should predict how much peptide that contour yields. Neither the phase-imaging front end nor the mass-based prediction of protein input has been tested in archival spatial proteomics. The research file's search found no work connecting unit-modulus optimization to this domain.
Why this intersection, and not the others
The research file suggests several other bridges. We rejected them because the unit-modulus structure would be imposed rather than found. Mass-spectral intensities are real and positive, so recasting DIA fragment unmixing as a problem on the unit circle has no physical basis. Library-free deep deconvolution such as Dear-DIA XMBD, a variational autoencoder that groups co-eluting fragments, is a far more natural tool there.
Coding ion-injection gates with radar-style unimodular sequences, in the spirit of Accelerated time-of-flight mass spectrometry, would require modifying the instrument. It also has nothing to do with archival tissue in particular.
The optical phase of a thin transparent specimen is different: it is genuinely unit-modulus. Archival tissue is a setting where going stain-free has a concrete payoff. Deep Visual Proteomics motivates the work in two ways. It shows that image-guided cutting of 17-year-old melanoma blocks is feasible. It also acknowledges that classifying cells from a 2D image while cutting a 3D slab mixes material from neighboring compartments.
Extraction work such as Davidson et al. shows how much chemistry, including heat and SDS, deoxycholate and Triton cocktails, is needed to recover proteins from fixed tissue. It is plausible, though not shown in the file, that exogenous stains and antibodies interfere further. This is a hypothesis we test, not a premise.
Research Program
Forward model and optimization. Fourier ptychography, the imaging method we build on, illuminates the sample from many LED angles and records one low-resolution intensity image per angle. We model each measurement as follows:
$$\min_{x,\;u_k}\ \sum_{k}\big\|A_k x - \sqrt{I_k}\odot u_k\big\|_2^2 \quad \text{s.t. } |x_i|=1,\ |u_{k,j}|=1.$$
The symbols are defined as follows. $x$ is the vector of complex sample transmissions, one per pixel. $A_k$ is the known linear optical operator for LED angle $k$: a tilt, a pupil cutoff and propagation. $I_k$ is the recorded intensity image for that angle. $u_k$ holds the unknown phases of the field at the camera for that measurement.
This split is chosen so that each subproblem matches a studied method. With $x$ fixed, the optimal $u_k$ is the closed-form phase extraction $u_k=\exp(j\arg(A_k x))$. This is the same element-wise step as the phase-extraction update in Yu et al.'s alternating minimization for hybrid precoding, which the file reports matches or beats Orthogonal Matching Pursuit in spectral efficiency and speed.
With $u_k$ fixed, the problem in $x$ is exactly unit-modulus least squares. We solve it with the backtracking and adaptive-restart accelerated projected gradient methods of Zheng et al., who prove local linear convergence when the reduced Riemannian Hessian is positive definite. We take the method from that paper, but its guarantee covers only the inner subproblem, not the outer alternation. We will measure empirically whether the Hessian condition holds on real tissue operators.
Handling pigment and residual absorption. Tissue is not perfectly transparent. Melanin in melanoma and hemosiderin both absorb, which breaks the pure-phase prior. Rather than ignore this, we relax the constraint to the convex hull $|x_i|\le 1$ and add a negative quadratic penalty weighted by a spatial map $\lambda_i$. This is the exact-penalization device of Extreme Point Pursuit, which proves that a large enough finite penalty drives the solution onto the unit circle. We also borrow its homotopy schedule, which raises the penalty weight gradually from zero, as a way to avoid bad early minima.
Pixels where the data resist being pushed to the circle keep $\lambda_i$ small and are allowed to absorb. That set of pixels is itself a pigment map. Here Extreme Point Pursuit constrains our design as much as it supplies it: its guarantees assume a smooth objective, and any we claim will be restricted to that setting.
From phase to proteomic decisions. We segment nuclei and cells from the phase images, either by retraining the segmentation stage of the pipeline used in Deep Visual Proteomics or by fine-tuning it on phase inputs. For each contour we integrate the phase over its area to get optical path length, which by the standard refractive-increment relation is proportional to dry mass.
That per-contour mass is used three ways. We report it as a predictor of peptide yield. We use it to pool contours to a fixed mass rather than a fixed cell count, which should reduce missing values at low input. We flag contours whose mass is too high for a single cell layer as likely overlaps.
This last use is deliberately modest. Phase integrates through the section's thickness, so it can flag stacking but cannot resolve depth. We do not claim to solve the 3D admixing problem.
Baselines
The naive reference point is the Deep Visual Proteomics workflow as published: a fluorescent or chromogenic stain, deep-learning nucleus segmentation, then microdissection and diaPASEF acquisition. Pools are defined by cell count and contour area, with no mass estimate. It is the strongest existing method for this question because it is the only system in the file that goes from archival image to cell-type proteome. On FFPE melanoma it reports more than 5,000 proteins from as few as 100 cells, with quantitative reproducibility of Pearson r = 0.96 (Mund et al.).
We will run it on adjacent serial sections from the same blocks. It will use the same instrument, the same extraction buffer, the same number of contours and the same DIA search settings. Where the file gives no number for our exact tissues, we compare against our own rerun rather than the published figure.
The stain-free arm beats this baseline if two things hold together. It must identify at least as many proteins per 100-cell pool with equal or better between-replicate correlation. And its phase-based segmentation must match stain-based segmentation within a pre-registered margin of nucleus-level F1 score on sections stained after imaging.
The second baseline is an unconstrained complex reconstruction of the same LED data. It uses the identical $A_k$ operators, loss, iteration budget and step-size search, but drops $|x_i|=1$. This is the ablation that isolates our key ingredient. The constraint earns its place only if it reaches a target phase error with clearly fewer LED angles.
The third baseline is a linearized weak-phase reconstruction, such as differential phase contrast from four half-pupil images. It is fast and needs no non-convex optimization, and it too gets the same raw images.
The fourth baseline, for the dry-mass claim, is contour area. Area is the free predictor that Deep Visual Proteomics already has. The mass estimate beats it only if it explains significantly more of the variance in summed precursor intensity per pool.
Why not just…?
Why not just keep the stain, as Deep Visual Proteomics does? It works, and 5,000 proteins from 100 cells is a high bar. Our proposal has to earn its place by showing that the stain costs something measurable. The file contains no direct measurement of stain-induced loss in FFPE proteomics, so this is an open question rather than an established fact.
The stained-versus-unstained serial-section comparison settles it. If the stained arm loses nothing and segments better, the stain-free argument collapses. The contribution then reduces to the dry-mass estimate, which can be computed from phase images of stained sections too, at the cost of corrupted phase in absorbing regions. We will report that outcome honestly.
Why not just use an unconstrained or linearized phase reconstruction? Both are real threats. For truly thin, weak samples, a linearized reconstruction may already be good enough. The unit-modulus prior helps most when measurements are few and phase shifts are large, as in thick or dense regions such as tumor nests where the weak-phase approximation fails.
The angle-sweep experiment below decides this. If unconstrained or linearized reconstruction matches ours at every tested angle count, we drop the claim that the constraint matters. The project then continues as a phase-guided microdissection study using the simplest adequate reconstruction.
Why not just predict a virtual stain from brightfield images of unstained tissue with a neural network? Such a network can give good segmentation, but it produces no physically calibrated quantity. It cannot yield dry mass, and it gives no pigment map with a principled meaning. Its failures on unseen tissue types also go undetected. We include it as a secondary segmentation comparison only. A virtual stain cannot compete on the mass-prediction endpoint, which is the part of the proposal most specific to proteomics.
Proposed Experiments
Phantoms and ground truth. We first validate reconstruction on phase targets of known optical thickness and on cultured-cell slides with independently measured dry mass. We report phase error against the number of LED angles and against the inner-solver choice: fixed-step projected gradient, backtracking, or adaptive-restart acceleration. We also check whether observed convergence is linear as Zheng et al. predict. No reported result was found on Lacuna for unit-modulus phase reconstruction of FFPE tissue, so these numbers set the first reference.
Archival tissue. We then image unstained sections from two FFPE settings the file documents. The first is primary melanoma of the kind used in Deep Visual Proteomics, which stresses the pigment relaxation. The second is gastric signet ring cell carcinoma blocks like those profiled by Jin et al., whose 7,322-protein library and four proteomic subtypes provide a biological sanity check. We will test whether stain-free pools recover subtype-discriminating proteins such as PRDX2 and DDX27 at the same rank as stained pools.
Segmentation is scored against expert annotation on the same sections after they are stained post-imaging. Proteomic endpoints are measured on serial sections: proteins identified, missing-value rate, and replicate Pearson correlation.
Decisive checks. Three results decide the project. First, the angle sweep must show the unit-modulus reconstruction reaching the phantom error target with substantially fewer exposures than the unconstrained ablation. Without that, the optimization contribution is void.
Second, in matched serial sections, stain-free pools must be non-inferior to stained pools in protein count and reproducibility. Third, per-contour phase-derived mass must predict per-pool peptide signal better than area does. Mass-normalized pooling must also reduce missing values at 100-cell input.
If only the third check passes, the result is still useful. Deep Visual Proteomics could then add a cheap mass measurement to its existing stained pipeline. The note would report that outcome as the finding, rather than stretching it into a stain-free claim.
References
- Deep Visual Proteomics defines single-cell identity and heterogeneity
- Dear-DIA XMBD : Deep Autoencoder Enables Deconvolution of Data-Independent Acquisition Proteomics
- Accelerated Time-of-Flight Mass Spectrometry
- Inter-Regional Proteomic Profiling of the Human Brain Using an Optimized Protein Extraction Method from Formalin-Fixed Tissue to Identify Signaling Pathways
- Alternating Minimization Algorithms for Hybrid Precoding in Millimeter Wave MIMO Systems
- On Local Linear Convergence of Projected Gradient Descent for Unit-Modulus Least Squares
- Extreme Point Pursuit - Part I: A Framework for Constant Modulus Optimization
- A comprehensive proteomic analysis uncovers novel molecular subtypes of gastric signet ring cell carcinoma: Identification of potential prognostic biomarkers and therapeutic targets