Structured Detection Microscope (SDM) Overview

• SDM is a class of microscopy methods that engineer the detector path—using phase elements, scanning, or diffractive coding—to extract more information per acquisition.
• Key architectures include remote scanning for increased spatial bandwidth, PSF engineering for super-resolution via spatial-mode demultiplexing, and multiplexed sensor arrays for large-scale compressive imaging.
• Unlike structured illumination, SDM strategically modulates detection, offering benefits like high throughput, precise optical sectioning, and reduced mechanical motion for motion-sensitive samples.

Searching arXiv for the cited SDM-related papers and closely related terminology. Structured Detection Microscope (SDM) denotes a class of microscopy strategies in which the detection path is deliberately structured so that information is redistributed, multiplexed, or computationally separable at the detector. Recent arXiv papers use closely related structured-detection language for several distinct architectures: detection-side remote scanning to increase the spatial bandwidth product of a camera-limited microscope, point-spread-function engineering for spatial-mode demultiplexing and deep super-resolution, structured detection in image scanning microscopy for simultaneous super-resolution and optical sectioning, and diffractive multiplexing across a gappy multi-sensor array for snapshot large-scale imaging (Kersuzan et al., 27 Mar 2025, Booth et al., 1 Apr 2026, Zunino et al., 2024, Zhou et al., 19 Jul 2025). In all of these cases, the central intervention is applied on the detection side rather than by temporally patterned illumination.

1. Terminology and conceptual scope

The term is not used uniformly. In "A detection module to increase the spatial bandwidth product of a microscope" (Kersuzan et al., 27 Mar 2025), the authors do not use the name “Structured Detection Microscope (SDM)”; they describe a “detection module” that inserts a scanning element in the detection path of a standard microscope to acquire sequential sub-images within the microscope objective’s full field without any sample motion. In "Structured Detection for Simultaneous Super-Resolution and Optical Sectioning in Laser Scanning Microscopy" (Zunino et al., 2024), the authors likewise do not name a microscope “Structured Detection Microscope”; the instrument is a laser scanning microscope in an ISM configuration with a SPAD array, and the method is referred to as “s^2ISM.” By contrast, "Structured detection microscopy" (Booth et al., 1 Apr 2026) explicitly uses SDM for a PSF-engineering approach to spatial-mode demultiplexing in fluorescence microscopy.

This usage indicates that SDM is best understood as a detection-side design principle rather than a single canonical instrument. What unifies the literature is the decision to structure detection by means of a pupil-plane phase element, a detector-array forward model, a detection-side scan, or a diffractive code spread across a sensor array, with the aim of increasing the information extracted per acquisition.

Variant	Detection structure	Reported function
Detection-side remote scanning	Scanner in the Fourier plane of the detection arm	Increased camera-accessible SBP without sample motion
PSF-engineered SDM	Quadrant phase plate in the Fourier plane	Spatial-mode demultiplexing and deep super-resolution
s^2ISM	SPAD array plus maximum-likelihood inversion	Simultaneous super-resolution and optical sectioning
Diffractive multiplexed SDM	DOE plus 48-sensor “super-sensor”	Snapshot gap filling and large-scale compressive imaging

A recurrent misconception is that structured detection is simply another name for structured illumination microscopy (SIM). The cited works distinguish the two explicitly: SIM modulates illumination to shift spatial frequencies or to create temporal illumination diversity, whereas structured detection keeps illumination conventional or widefield and places the coding, multiplexing, or discrimination burden in detection (Booth et al., 1 Apr 2026, Zhou et al., 19 Jul 2025).

2. Shared detection-side principle

Across the literature, structured detection modifies the microscope forward model by altering how emitted or transmitted light is accepted, shaped, indexed, or scanned after the specimen. In the PSF-engineering formulation, the pupil function $P(k_x,k_y)$ is multiplied by a detection mask $M(k_x,k_y)$, giving an intensity PSF

$$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$

so that estimation performance is changed by design rather than by post hoc denoising (Booth et al., 1 Apr 2026). In the diffractive multiplexing formulation, a DOE phase $\phi(u,v)$ is inserted in the pupil plane and the incoherent PSF becomes

$$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$

with the measurements modeled as

$y = M\,(h * x) + n,$

where $M$ is an erasure mask for inactive sensor regions (Zhou et al., 19 Jul 2025).

In structured detection for ISM, the critical information channel is not a shaped image-plane PSF alone but the detector-array “fingerprint.” For detector element $i$ at position $r_{di}$, the forward model for a single focal-plane acquisition is written as

$$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$

with separate in-focus and out-of-focus planes estimated by maximum-likelihood inversion (Zunino et al., 2024). In detection-side remote scanning, by contrast, the scan element is placed in a Fourier plane, so mirror tilt produces an image translation rather than a geometric distortion:

$M(k_x,k_y)$0

This allows the camera’s fixed pixel budget to be reused over multiple non-overlapping tiles without moving the specimen (Kersuzan et al., 27 Mar 2025).

These formulations have different objectives—super-resolution, optical sectioning, field expansion, or snapshot throughput—but they share a common logic: the detector does not merely record a passive image; it participates in a deliberately designed measurement operator.

3. Detection-side remote scanning for spatial bandwidth product expansion

The remote-scanning implementation reported in "A detection module to increase the spatial bandwidth product of a microscope" (Kersuzan et al., 27 Mar 2025) is a detection module that attaches to the output port of a commercial inverted microscope. The optical layout splits the microscope’s image-forming light into two detection channels via a non-polarizing beamsplitter. Channel 1 provides a wide field with moderate sampling: L1 ($M(k_x,k_y)$1 mm) and L2 ($M(k_x,k_y)$2 mm) form an afocal system of magnification $M(k_x,k_y)$3, relaying to a Thorlabs Zelux scientific CMOS camera and yielding an apparent pixel size of $M(k_x,k_y)$4m at the sample with a 10×, NA 0.5 Nikon CFI Super Fluor objective. The measured field is about $M(k_x,k_y)$5 mm $M(k_x,k_y)$6 $M(k_x,k_y)$7 mm. Channel 2 provides a tunable ROI with fine sampling: after the beamsplitter, light reflects from a motorized 2D steering mirror (Optotune MR-15-30) placed in a Fourier plane and is then imaged by L3 ($M(k_x,k_y)$8 mm) onto a second Zelux camera, with each pixel mapping to $M(k_x,k_y)$9 nm at the sample and a typical single-frame FOV of $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$0m $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$1 $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$2m.

The mathematical motivation is the spatial bandwidth product (SBP). For incoherent imaging,

$$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$3

with $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$4 and Nyquist sampling constraint $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$5. The paper also invokes an objective-centric estimate,

$$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$6

For a 10×/0.5 objective at $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$7 nm with a nominal $$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$8 mm$$h(x,y)=\left|\mathcal{F}\{ P(k_x,k_y)\, M(k_x,k_y) \}\right|^2,$$9 FOV, this gives approximately $\phi(u,v)$0 megapixels. The detection module nevertheless collected over a larger effective diameter, up to approximately $\phi(u,v)$1 mm, which explains why the stitched image contains approximately $\phi(u,v)$2 Mpixels of information while maintaining Nyquist sampling near the central field.

The essential scaling law is

$\phi(u,v)$3

where $\phi(u,v)$4 is the number of non-overlapping scan positions and $\phi(u,v)$5 is the camera format per frame. In the reported bright-field scan, $\phi(u,v)$6 for a $\phi(u,v)$7 mirror sequence, producing a mosaic of approximately $\phi(u,v)$8 pixels, approximately $\phi(u,v)$9 Mp total, of which an approximately $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$0-pixel-diameter disk, approximately $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$1 Mp, contains objective-transmitted information. Full-FOV bright-field scans complete in approximately $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$2 s; the mirror’s settling time is approximately $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$3 ms and is small compared with typical exposure times. In epifluorescence, single-frame exposures of $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$4 ms yield a full-FOV scan in approximately $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$5 s.

Experimentally, the method operates in bright-field, phase contrast, and epifluorescence. The measured lateral resolution is relatively uniform within the objective’s nominal field number, approaching the theoretical approximately $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$6m; at the outer field, coma degrades resolution to $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$7–$$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$8m. Collection efficiency is homogeneous over more than $$h(x,y)=\left|\mathcal{F}^{-1}\{\,P(u,v)\,e^{i\phi(u,v)}\,\}\right|^2,$$9 mm diameter before roll-off, and because the steering mirror is in the Fourier plane, mirror tilts produce pure translations in the image plane without geometric distortion. Registration uses a single pre-calibration from mirror angles to tile coordinates and the Fiji stitching plugin invoked from Python via PyImageJ. For sparse ROI time-lapse, the system can acquire manually selected positions only; in one example, monitoring 5 ROIs every 4 min for 3 h produced approximately $y = M\,(h * x) + n,$0 MB of data, versus approximately $y = M\,(h * x) + n,$1 GB if the entire FOV had been captured at the same cadence.

This architecture is presented as particularly advantageous for motion-sensitive samples. The comparison to stage scanning is not given as a head-to-head throughput benchmark, but the reported advantages are explicit: no mechanical motion of the specimen, ms-level mirror settling rather than tens–hundreds of ms stage settling, reduced registration errors because the tiles are pure translations, and the ability to scan sparse ROIs without collecting large, mostly irrelevant datasets.

4. PSF-engineered SDM as spatial-mode demultiplexing

In "Structured detection microscopy" (Booth et al., 1 Apr 2026), SDM is a camera-based, PSF-engineering approach to spatial-mode demultiplexing for fluorescence microscopy. The stated objective is to avoid nonlinear saturation and stochastic emitter switching while increasing the Fisher information for estimating sub-diffraction separations between incoherent emitters. Rather than using mode-sorting optics, the method reshapes the conventional near-Gaussian PSF into a four-lobed “split-Gaussian” mode that approximates a 2D Hermite–Gaussian TEM11 mode. A quadrant phase plate in the Fourier plane flips the phase by $y = M\,(h * x) + n,$2 in two diagonally opposite quadrants, creating zeros near the centroid and moving separation-sensitive signal away from high shot-noise PSF peaks.

For two incoherent emitters of equal brightness with separation components $y = M\,(h * x) + n,$3, the per-pixel mean count under conventional Gaussian imaging is modeled as

$y = M\,(h * x) + n,$4

whereas the SDM split-Gaussian model introduces the multiplicative $y = M\,(h * x) + n,$5 structure characteristic of the derivative-like mode. Photon counts are modeled as Poisson with additive white background, and inference proceeds through a log-likelihood and Fisher information matrix. The key claim is that for small separations $y = M\,(h * x) + n,$6, SDM concentrates $y = M\,(h * x) + n,$7 near low-intensity PSF regions, so Fisher information degrades linearly as $y = M\,(h * x) + n,$8, whereas conventional imaging degrades quadratically. The theoretical transition where SDM overtakes conventional imaging is predicted at $y = M\,(h * x) + n,$9, approximately $M$0 nm for visible wavelengths.

The implementation is a high-NA TIRF microscope. The platform is a commercial 3i Marianas TIRF microscope with an oil-immersion 100×, NA $M$1 objective, $M$2 nm, and $M$3 nm. The theoretical Rayleigh diffraction limit at $M$4 nm is approximately $M$5 nm, and the measured diffraction limit is approximately $M$6 nm. The quadrant phase plate is fabricated from a Union Optic zero-order half-waveplate, diced and reassembled to produce the required $M$7 phase flips. Relay optics consist of L1 ($M$8 mm), L2 ($M$9 mm), and a telescope L3 ($i$0 mm), L4 ($i$1 mm) to magnify the PSF by 8×. Detection uses an iXon Ultra 897 EMCCD with QE approximately $i$2, cooled to $i$3C, EM gain approximately 300×, and electronic noise approximately $i$4 s$i$5 per pixel; the effective pixel size on the detection camera is approximately $i$6 nm.

The reported sample is DNA nanorulers of 50, 120, and 180 nm length, with each end labeled by a cluster of 20 Alexa Fluor 488 fluorophores. Typical exposure is 500 ms per frame, with photobleaching over approximately 3 s and photon counts of order $i$7 per image on average. SDM resolves separations as small as 50 nm and achieves RMSE resolutions of 39 nm for 50 nm rulers, 28 nm for 120 nm rulers, and 18 nm for 180 nm rulers; conventional imaging gives 64 nm, 53 nm, and 27 nm, respectively. The paper reports resolution improvement factors of 1.6×, 1.9×, and 1.5×, and characterizes the 39 nm result as approximately fivefold below the diffraction limit. At 50 nm, the theoretical CRLB is approximately 31 nm for SDM, with experimental performance limited by ruler length uncertainty, residual bias, background nonuniformity, pixelation artifacts, finite FOV, and model–PSF mismatch.

Reconstruction proceeds by cropping ROIs around individual emitter pairs, estimating background from image corners, fitting the PSF model to determine centroid and widths, building a separation grid super-sampled at 1/18 pixel spacing, computing per-photon likelihoods, forming a posterior over $i$8, and then applying a calibrated sigmoid-based bias correction. The scope is correspondingly narrow and explicit: the present implementation assumes two incoherent emitters of similar brightness, sparse fields, known emitter number per ROI, and 2D separations near the coverslip in TIRF imaging.

5. Detector-array and sensor-array realizations

Structured detection in scanning microscopy is developed in "Structured Detection for Simultaneous Super-Resolution and Optical Sectioning in Laser Scanning Microscopy" (Zunino et al., 2024). The hardware is a laser scanning microscope in an image-scanning configuration with a 7×7 SPAD array, of which the inner 5×5 are read. The authors exploit the fact that the array’s micro-images encode axial information in the distribution of detected photons. For a point emitter, the detector-plane integrated signal defines a “fingerprint” that broadens with defocus: central detector elements dominate in focus, outer elements contribute relatively more out of focus. The inversion therefore estimates two images from a single 2D scan: an in-focus super-resolved image and an out-of-focus background image, with photon counts modeled as Poisson and recovered by a Richardson–Lucy-type multiplicative iteration under nonnegativity. The method reduces to ISM multi-image deconvolution when the number of axial planes is set to one.

The reported gains are quantitative. On a fluorescent line target at $i$9 nm, the dip-visibility cut-off spacing improves from 210 nm for open-pinhole confocal to 150 nm with structured detection. On a 3D stair of rings with $r_{di}$0 nm, the optical sectioning function FWHM is reduced by roughly $r_{di}$1 versus the open-pinhole confocal. When the scan pixel size equals the projected detector pitch, $r_{di}$2, the method supports two-fold upsampling, relaxing Nyquist’s criterion by a factor of two; this was verified experimentally with nuclear pore data, reporting SSIM up to 0.98 and median approximately 0.87. The same forward model was generalized to time-resolved FLIM using measured per-element IRFs, improving axial cross-talk suppression and tightening phasor clouds in mixed-lifetime specimens.

A distinct large-scale realization appears in "Large-scale compressive microscopy via diffractive multiplexing across a sensor array" (Zhou et al., 19 Jul 2025). Here the optical train is approximately 4f, a DOE is placed at the pupil plane, and the output is recorded by a 6×8 grid of 48 monochrome sensors treated as one contiguous “super-sensor” with erasures in the gaps between active dies. The DOE engineers a sparse multi-impulse PSF with $r_{di}$3 diffraction-limited spots over a footprint of approximately $r_{di}$4 mm $r_{di}$5 $r_{di}$6 mm, larger than the largest gap. The design goals are pan-visibility, minimal translational ambiguity, narrow extent, sparsity, and robustness to scaling errors. Reconstruction solves

$r_{di}$7

with non-stochastic gradient descent, backtracking, and non-negativity projection.

This system is optimized for spatiotemporal throughput rather than sub-diffraction resolution. The raw array contains 48 sensors, each 3136×4096 pixels with 1.1 $r_{di}$8m pitch and 8-bit readout. The total data bandwidth is 6.17 GB/s; full-array raw acquisition at 0.617 gigapixels/frame is bandwidth-limited to approximately 10 fps. In practice, 4×4 binning enables 120 fps. Darkfield reconstructions cover approximately $r_{di}$9 mm$$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$0, approximately $$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$1 cm$$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$2, at approximately $$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$3m resolution and approximately 25.2 billion pixels per second spatiotemporal throughput. Fluorescence reconstructions cover approximately $$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$4 mm$$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$5, and the paper demonstrates pharyngeal GCaMP8f dynamics of 12 worms at 120 fps as well as structural darkfield imaging of dozens of freely moving C. elegans over 15 s. The system is described as calibration-free in the sense that no empirical per-pixel PSF measurement is required; DOE phase, sensor geometry, and distortion parameters determine the forward model.

Taken together, these two works show that structured detection is not tied to a single detector technology. In one case, the discriminative variable is the detector-array fingerprint under point scanning; in the other, it is the distributed PSF over a gappy super-sensor in a single snapshot.

6. Trade-offs, relations to adjacent methods, and current limits

The principal trade-offs differ across SDM variants. In detection-side remote scanning, the dominant constraints are the microscope objective’s off-axis aberrations, field curvature, vignetting, photon splitting at the beamsplitter, and scanner dynamics. Resolution remains near diffraction-limited within the nominal field number but degrades to $$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$6–$$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$7m at the outer extended field because of coma; collection efficiency is homogeneous over more than $$i(r \mid r_{di}) = \sum_k f_k(r_{di}) [o_k(r) * \hat h_k(r \mid r_{di})],$$8 mm diameter before rapid fall-off; and in low-light epifluorescence, the beamsplitter necessitates longer exposures such as 180 ms per tile (Kersuzan et al., 27 Mar 2025). In PSF-engineered super-resolution SDM, the key limitations are background sensitivity, aberrations and misalignment, drift and pixelation artifacts, the need for bias correction, and the present restriction to sparse ROIs containing two incoherent emitters of similar brightness (Booth et al., 1 Apr 2026). In s^2ISM, over-iteration amplifies noise, strongly asymmetric or space-variant aberrations degrade PSF fidelity, and very scattering samples can distort the fingerprint, although two-photon excitation improves penetration and sectioning (Zunino et al., 2024). In diffractive multiplexed SDM, dense scenes, photon starvation, vignetting, unmodeled aberrations, and violation of sparsity or compressibility assumptions degrade reconstruction quality, and the inverse problem imposes a nontrivial computational cost (Zhou et al., 19 Jul 2025).

The relation to neighboring methods is explicit in the cited literature. Structured detection is not SIM: SIM varies illumination patterns temporally, whereas these SDM variants encode information in detection. It is also distinct from STED and PALM/STORM. The PSF-engineered SDM work specifically presents itself as a route to deep super-resolution without saturation or stochasticity, with photon doses similar to standard widefield/TIRF for sparse emitter pairs. The remote-scanning work contrasts its detection-side scan with motorized stage tiling; the compressive multi-sensor work contrasts its snapshot coding with multi-shot Fourier ptychography and with diffuser-like coded systems that use more diffuse PSFs; and s^2ISM positions structured detection as a generalization of ISM multi-image deconvolution that additionally performs optical sectioning.

A second misconception is that SDM necessarily implies super-resolution below the diffraction limit. The literature shows otherwise. The PSF-engineered SDM of (Booth et al., 1 Apr 2026) achieves far sub-wavelength localization of pair separations; s^2ISM achieves lateral super-resolution together with optical sectioning; the remote-scanning module of (Kersuzan et al., 27 Mar 2025) does not improve the objective’s intrinsic resolution but increases the accessible SBP by reusing the camera across many translated sub-fields; and the diffractive multi-sensor system of (Zhou et al., 19 Jul 2025) is optimized for gigapixel-scale throughput and large FOV rather than nanometric resolution. The commonality is structured detection, not a single performance regime.

The papers also identify concrete extension paths. The remote-scanning module could reduce overhead further with faster galvanometric scanners or tighter timing with hardware TTL triggering. The PSF-engineered SDM work states that adaptive optics phase arrays could generalize SDM/SPADE to arbitrary distributions and potentially to 3D. The s^2ISM framework already extends to FLIM and to nonlinear excitation. The diffractive multi-sensor system could incorporate richer priors, learned reconstructions, or more elaborate spatially varying PSF models to handle dense scenes and peripheral aberrations. This suggests that SDM is less a fixed microscope category than an expanding design space in which detection itself becomes the primary site of optical coding and inference.