Ultra-high-speed line-scan Raman imaging

Bei Li; Francesco Masia; Hao Peng; Lindong Shang; Peng Liang; Qingyi Wu; Wolfgang Langbein; Xusheng Tang; Yue Qu; Yuntong Wang

REVIEW 2 major objections 5 minor 10 references

Ultra-high-speed line-scan Raman imaging

T0 review · 2 major / 5 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A low-cost CMOS line-scan Raman system images full spectra at up to 80 kHz, two orders faster than prior line-scan methods.

desk verdict Solid instrumentation paper that delivers a real, quantified 100× speed jump in full-spectrum line-scan Raman by swapping a modern global-shutter CMOS for the usual CCD bottleneck. read the letter →

arxiv 2607.04877 v1 pith:TLJENWLV submitted 2026-07-06 physics.optics cond-mat.mtrl-sci

Qingyi Wu , Xusheng Tang , Francesco Masia , Peng Liang , Hao Peng , Lindong Shang , Yuntong Wang , Yue Qu

show 2 more authors

Wolfgang Langbein Bei Li

This is my paper · ORCID

classification physics.opticscond-mat.mtrl-sci PACS 42.62.Fi07.60.Rd87.64.kp

keywords line-scanRamanCMOSdetectorhyperspectralimagingmicroplasticsFSC3factorizationhigh-speedspectroscopypharmaceuticaltabletsbiologicalcells

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper shows that modern machine-vision CMOS cameras, which read out while exposing under a global shutter, remove the long-standing speed bottleneck of CCD detectors in line-scan Raman imaging. Combined with a high-efficiency transmission-grating spectrometer and continuous sample motion, the instrument records full Raman spectra at rates up to 80 kHz (and hardware-supported rates near 190 kHz). That is two orders of magnitude faster than conventional line-scan Raman and three-to-four orders faster than point-scan confocal Raman. The authors demonstrate the system on microplastics, dried cells and pharmaceutical tablets, recovering chemically distinct Raman and fluorescence maps in seconds rather than minutes or hours. Because the hardware is inexpensive and the detection remains near the shot-noise limit for the demonstrated samples, the method opens practical high-throughput chemical imaging for environmental, biological and pharmaceutical work.

What carries the argument

Overlapping global-shutter CMOS readout during continuous stage motion, paired with FSC3 non-negative matrix factorization that jointly extracts Raman and fluorescence components by smoothness nudging rather than per-spectrum background fitting.

What would settle it

Acquire the same microplastic or cell field at the paper’s fastest settings and at a ten-fold longer exposure (or with a lower-noise camera) and show that the FSC3 component maps and spectra change systematically rather than remaining stable within shot noise.

Watch

Extended reading notes

Core claim

Replacing cooled CCD detectors with a low-read-noise global-shutter CMOS camera and a transmission-grating spectrometer allows spontaneous line-scan Raman imaging at spectral rates of 80 kHz (hardware up to ~191 kHz), two orders of magnitude faster than previous line-scan instruments and up to four orders faster than point scanning, while remaining near shot-noise limited for the samples shown.

Load-bearing premise

That the authors’ unpublished smoothness-nudging version of FSC3 recovers chemically faithful Raman and fluorescence maps even when each pixel contains only a few to tens of photoelectrons.

Editorial extensions

If this is right

Megapixel spontaneous Raman images can be collected in roughly 12 seconds under continuous motion.
Higher laser power or sub-electron-read-noise cameras would push shot-noise-limited spectral rates another order of magnitude higher.
The same hardware is immediately usable for ultrafast spatially-offset Raman tomography of tablets and other turbid samples.
GPU-based real-time FSC3 could display chemical concentration maps during acquisition rather than after offline processing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the camera and optics are already commodity items, the cost barrier that has kept line-scan Raman out of routine environmental or industrial labs is largely removed.
The demonstrated speed may finally make spontaneous full-spectrum Raman competitive with coherent techniques for dynamic processes that last tens of seconds.
The same continuous-motion line-scan architecture could be ported to other weak-signal spectroscopies (e.g., fluorescence lifetime or Brillouin) that currently rely on slow CCD detectors.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The headline result is real: continuous-motion line-scan with a cheap machine-vision CMOS (global shutter, overlapping readout) plus a high-efficiency VPH spectrometer reaches 80 kHz spectral rate on actual samples, two orders of magnitude faster than prior line-scan work (including the authors’ own CCD system) and near shot-noise limited for the microplastic data. That is the new part. Everything else—Powell-lens line generation, dual-NF confocal slits, FSC3 post-processing—is incremental or already published.

What they do well is the engineering detail. Optical layout, timing diagram (40 µs overhead), stage velocities, power levels, photoelectron counts, and the raw-data link are all there. You can reproduce the hardware performance from the text. The microplastic, cell, and tablet examples show the system works on real samples without obvious photodamage at the stated powers. The noise budget for Fig. 2 is honest: they correctly note that 2 ms exposures become read-noise limited and that higher laser power or sub-1 e- sCMOS would push the limit further.

Soft spots are secondary and proportional. The improved FSC3 smoothness-nudging that replaces conventional fluorescence fitting is unpublished and therefore a black box for the chemical maps; at a few-to-tens of pe per pixel it could introduce artefacts. That does not touch the spectral-rate claim itself. The aluminium-substrate enhancement discussion is a bit hand-wavy, and the tablet crystal-orientation interpretation is qualitative. None of these are load-bearing for the central result.

This is for groups who already do spontaneous Raman imaging and want higher throughput without going nonlinear. It is not a new scientific domain, but it is a practical step-change that will be cited. The math, data, and citations look solid; no circularity. I would send it to referees without hesitation.

Referee Report

2 major / 5 minor

Summary. The manuscript presents an ultra-fast line-scan Raman micro-spectroscopy system that replaces conventional CCD detectors with a low-cost, global-shutter machine-vision CMOS camera (Sony IMX392) and a high-efficiency volume-phase holographic transmission-grating spectrometer. Continuous stage motion synchronized with overlapping frame acquisition yields spectral rates up to 80 kHz (478 Hz frames of 400 spectra), two orders of magnitude faster than prior line-scan Raman instruments and up to four orders faster than point-scan methods. Optical layout, power levels, exposure times, stage velocities, photoelectron counts, and noise budgets are fully specified. The system is demonstrated on microplastics (15 s for a 1000 × 1200 pixel map), dried HepG2 cells on aluminium-coated and CaF2 substrates, and acetaminophen tablets. Hyperspectral data are factorized by an improved FSC3 non-negative matrix factorization that separates Raman and fluorescence components without a prior background fit.

Significance. If the reported hardware performance holds, the work removes a long-standing detector bottleneck in spontaneous Raman imaging and makes megapixel chemical maps practical on a seconds-to-minutes timescale with inexpensive components. The combination of continuous-motion acquisition, global shutter, and high-étendue transmission spectrometer is a clear engineering advance over earlier line-scan systems that used cooled CCDs. The open raw-data link and detailed noise analysis further strengthen the contribution. The chemical-map interpretation rests on an unpublished FSC3 smoothness-nudging procedure, but this is secondary to the primary speed claim.

major comments (2)

Results §2 (cells) and Materials and Methods §3: the improved FSC3 procedure that replaces conventional fluorescence fitting by iterative smoothness-nudging is described only qualitatively and is stated to be published elsewhere. Because the microplastic and cell data operate at only a few to tens of photoelectrons per pixel, the chemical assignments (PP/PVC/PS, lipid vs nucleic-acid components) rest on this unpublished algorithm. A concise description of the nudging rules, the number of free parameters, and a side-by-side comparison with a conventional background-subtracted NMF on at least one data set would make the chemical maps reproducible and would strengthen the secondary claim of “exceptional detection capabilities.”
Discussion and Results §1: the claim of “up to two orders of magnitude faster than traditional line-scan Raman” is supported by comparison with the authors’ own prior CCD work, but quantitative benchmarks against other recent line-scan or multifocal systems (e.g., those using EMCCDs or sCMOS) are missing. Adding a short table of spectral rates, power densities and SNR metrics for the closest published instruments would place the advance more rigorously.

minor comments (5)

Fig. 1 caption and Materials and Methods: several lens and filter part numbers are listed, but the effective numerical aperture and collection solid-angle calculation for the aluminized versus transparent substrates would benefit from an explicit formula or short derivation.
Results §1 (microplastics): the statement that detection is “at the limit of being shot-noise limited” is correct for the summed signal, yet the peak signals are only a few pe; a brief note on how spatial binning or spectral binning would alter this balance would help readers planning similar experiments.
Supplementary Fig. S9: the conversion factor between counts and photoelectrons is given as 0.084 pe/count for 12-bit data; clarifying whether this factor is applied uniformly to all presented spectra would avoid confusion.
Author list and affiliations: the repeated “State Key Laboratory” affiliation for the first three authors can be consolidated for clarity.
References: a few recent high-speed Raman or sCMOS-based systems (e.g., the 2024 miniaturized confocal system already cited) could be expanded slightly to acknowledge parallel detector developments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: experimental speed claim is a direct hardware measurement, not a derived prediction forced by inputs or self-citation.

full rationale

The paper's central claim (Raman spectral rates up to 80 kHz, two orders of magnitude faster than prior line-scan work) is obtained by arithmetic from measured camera frame rates (up to 478 Hz for 1920 imes400), exposure times (2.05–5 ms), continuous stage velocity, and the fixed 400 spectra per line; these quantities are reported from the instrument and raw-data archive, not fitted or defined in terms of the claimed rate. FSC3 (and its unpublished smoothness-nudging extension) is used only for post-hoc chemical-map visualization of the already-acquired hyperspectral cubes; earlier FSC3 papers by overlapping authors are cited for the base algorithm, but the speed metric itself does not depend on them, nor is any uniqueness theorem or ansatz imported to force the rate result. No free parameters are fitted to data and then re-presented as predictions, and no self-definitional loop exists between the hardware description and the reported spectral rate. The work is therefore self-contained against external benchmarks of frame rate and photoelectron counts.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper is an experimental instrumentation demonstration. Almost all numerical values (powers, exposures, stage speeds, camera noise, QE) are measured or taken from manufacturer data sheets rather than fitted free parameters. The only free choices that affect the central claim are the number of FSC3 components and the (unpublished) smoothness-nudging thresholds used to separate Raman from fluorescence; these are analysis parameters, not physical free parameters of the optical system.

free parameters (2)

number of FSC3 components = 6–8
Manually chosen (6 for microplastics, 7 for Al-substrate cells, 5–8 for tablets) to ‘retrieve all relevant chemical components while avoiding splitting’; different choices alter the reported concentration maps.
FSC3 smoothness-nudging thresholds
Unpublished parameters that decide which spectral components are forced toward smooth fluorescence versus sharp Raman; they directly control the separation shown in the figures.

assumptions (3)

domain assumption The Sony IMX392 CMOS sensor has 2.5 e- read noise and ~60 % QE in 540–680 nm under the stated gain and 8-bit readout conditions.
Taken from manufacturer data and used to claim near-shot-noise-limited detection; not re-measured in the paper.
domain assumption Continuous constant-velocity stage motion at velocity = frame-rate × step-size produces spatially undistorted line-scan images without the need for ultrafast positioning.
Stated in Results §1; assumed valid for the 0.5–500 µm s-1 speeds used.
ad hoc to paper FSC3 non-negative matrix factorization with the authors’ smoothness-nudging procedure recovers chemically interpretable Raman and fluorescence components from low-SNR hyperspectral data.
The improved algorithm is introduced here and details deferred; the chemical assignments in Figs. 2–5 rest on it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Ultra-high-speed line-scan Raman imaging." pith.science (2026). https://pith.science/paper/TLJENWLV

@misc{pith2026260704877,
  author       = {Pith},
  title        = {Pith review of: Ultra-high-speed line-scan Raman imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLJENWLV}},
  note         = {Machine review of arXiv:2607.04877}
}

read the original abstract

Raman spectroscopic imaging has emerged as a potent tool due to its non-invasive nature and capability for chemical composition analysis. Line-scan Raman spectroscopy accelerates imaging speed by two orders of magnitude compared to point detection Raman methods. However, further enhancements in imaging speed were constrained by the readout speed of typically used charge-coupled device (CCD) spectroscopic detectors. We developed an ultra-fast line-scan Raman imaging technique based on recently available complementary metal-oxide-semiconductor (CMOS) detectors with low cost and read noise, and fast readout during exposure combined with a global shutter. Employing a high-efficiency transmissiongrating imaging spectrometer, we demonstrate imaging speeds up to two orders of magnitude faster than traditional line scan Raman imaging techniques and up to four orders of magnitude faster than point scan Raman methods, achieving Raman imaging up to 80 kHz spectral rate. We demonstrate that this technology is applicable to a variety of samples, including microplastics, biological cells, and tablets, creating images in an extremely short time frame, showcasing exceptional detection capabilities and the ability to reveal detailed information.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references

[1]

Diagram of acquisition timing Fig

Data acquisition Fig.S1. Diagram of acquisition timing Fig. S1. illustrates the acquisition tim ing for the CMOS camera. The CMOS camera is equipped with a global shutter. To achieve higher frame rates, the camera employs overlapping acquisition, which involves exposing a new image while the sensor data from the previous image is being read out. This oper...
[2]

Hot pixel correction Fig. S2. Sketch of hot pixels correction method Fig. S2 shows the method used for hot pixels correction. An average of ten frames with 5 s exposure time in a dark environment is taken and a threshold is determined by summing the mode of the distribution and three times the standard deviation of the intensity values in this average ima...
[3]

Tilt and Distortion Correction Fig. S3. Schematic diagram of the tilt and distortion of the line. (a) shows a vertical line spot on the CMOS sensor, (b) shows the tilt of the line, and (c) shows a quadratic cushion distortion of the line, with the top and bottom bent in the same direction. The pixels in the camera are identified by their coordinates (𝑠, 𝑦...

1920
[4]

Spectral calibration Atomic spectra of neon-argon (Ne -Ar) and mercury (Hg) lamps were collected using the system, and the pixel positions of each characteristic peak were determined using a Gauss - Lorentz fitting method, 𝑓(𝑥; 𝐴, 𝜇, 𝜎, 𝛼), given by 𝑓(𝑥; 𝐴, 𝜇, 𝜎, 𝛼) = (1−𝛼)𝐴 𝜎𝑔√2𝜋 𝑒[−(𝑥−𝜇)2 2⁄ 𝜎𝑔2] + 𝛼𝐴 𝜋 [ 𝜎 (𝑥−𝜇)2+𝜎2] , where 𝑥 is the pixel number ; 𝐴 i...
[5]

𝑆𝑀(𝜆) denotes the image data after removal of offset and dark noise

Offset and dark noise removal The offset of a detector is the baseline signal generated by the detector circuitry under no - light conditions, and the dark noise of a detector aris ing mainly from electronic thermal activity or charge leakage in the absence of illumination , and is proportional to the exposure time. 𝑆𝑀(𝜆) denotes the image data after remo...
[6]

Spectral intensity calibration Using the 532 nm standard reference material NIST, the spectral image data was corrected by obtaining an image intensity correction coefficient matrix from the actual measured NIST spectral image with its standard intensity distribution. Fig. S8 shows the comparison of the results before and after intensity correction on the...

1920
[7]

Larger singular values correspond to more dominant features in the data, while small singular values are usually associated with noise

Spectral denoising Using Singular Value Decomposition (SVD), the spectral data matrix 𝐷 is decomposed into 𝐷 = 𝑈𝛴𝑉𝑇where the matrix 𝛴 is a diagonal matrix where the diagonal elements are singular values, the matrix 𝑈 is the 𝑀 × 𝑀 matrix containing the set of left singular vectors (basis vectors), and the matrix 𝑉 is the 𝑁 × 𝑁 matrix containing the set of ...
[8]

Comparison of the substrate spectrum with fused silica The spectral shape of fused quartz (also called fused silica) from 300 to 600/cm is shown in Fig. S10. T he component Fig.2f , reproduced in Fig.S11 bottom left, appears to be a broadened version of this shape. Since the signal is very weak (1 pe maximum), the spectral shape of the component is less d...

2017

Show all 10 references

[9]

Larger plots of the spectra shown in all figures 2,3,4,5 Fig. S11. Larger graphs of the spectra shown in Fig 2 of the main text, provided for clarity. Fig. S12. Larger graphs of the spectra shown in Fig 3 of the main text, provided for clarity. Fig. S13. Larger graphs of the s...
[10]

Data The original data for the results shown in Fig.2,3,4,5 are provided under the link https://1drv.ms/u/c/b3219e77e76a7d44/IQDM1KSuClV2R6G1_lDytjboAfxMJtuxizrzTJKVv pBf0XA. They contain the raw data as 8bit images as acquired, the SVD denoised data, giving the spectra (text)...

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.

[1] [1]

Diagram of acquisition timing Fig

Data acquisition Fig.S1. Diagram of acquisition timing Fig. S1. illustrates the acquisition tim ing for the CMOS camera. The CMOS camera is equipped with a global shutter. To achieve higher frame rates, the camera employs overlapping acquisition, which involves exposing a new image while the sensor data from the previous image is being read out. This oper...

[2] [2]

Hot pixel correction Fig. S2. Sketch of hot pixels correction method Fig. S2 shows the method used for hot pixels correction. An average of ten frames with 5 s exposure time in a dark environment is taken and a threshold is determined by summing the mode of the distribution and three times the standard deviation of the intensity values in this average ima...

[3] [3]

Tilt and Distortion Correction Fig. S3. Schematic diagram of the tilt and distortion of the line. (a) shows a vertical line spot on the CMOS sensor, (b) shows the tilt of the line, and (c) shows a quadratic cushion distortion of the line, with the top and bottom bent in the same direction. The pixels in the camera are identified by their coordinates (𝑠, 𝑦...

1920

[4] [4]

Spectral calibration Atomic spectra of neon-argon (Ne -Ar) and mercury (Hg) lamps were collected using the system, and the pixel positions of each characteristic peak were determined using a Gauss - Lorentz fitting method, 𝑓(𝑥; 𝐴, 𝜇, 𝜎, 𝛼), given by 𝑓(𝑥; 𝐴, 𝜇, 𝜎, 𝛼) = (1−𝛼)𝐴 𝜎𝑔√2𝜋 𝑒[−(𝑥−𝜇)2 2⁄ 𝜎𝑔2] + 𝛼𝐴 𝜋 [ 𝜎 (𝑥−𝜇)2+𝜎2] , where 𝑥 is the pixel number ; 𝐴 i...

[5] [5]

𝑆𝑀(𝜆) denotes the image data after removal of offset and dark noise

Offset and dark noise removal The offset of a detector is the baseline signal generated by the detector circuitry under no - light conditions, and the dark noise of a detector aris ing mainly from electronic thermal activity or charge leakage in the absence of illumination , and is proportional to the exposure time. 𝑆𝑀(𝜆) denotes the image data after remo...

[6] [6]

Spectral intensity calibration Using the 532 nm standard reference material NIST, the spectral image data was corrected by obtaining an image intensity correction coefficient matrix from the actual measured NIST spectral image with its standard intensity distribution. Fig. S8 shows the comparison of the results before and after intensity correction on the...

1920

[7] [7]

Larger singular values correspond to more dominant features in the data, while small singular values are usually associated with noise

Spectral denoising Using Singular Value Decomposition (SVD), the spectral data matrix 𝐷 is decomposed into 𝐷 = 𝑈𝛴𝑉𝑇where the matrix 𝛴 is a diagonal matrix where the diagonal elements are singular values, the matrix 𝑈 is the 𝑀 × 𝑀 matrix containing the set of left singular vectors (basis vectors), and the matrix 𝑉 is the 𝑁 × 𝑁 matrix containing the set of ...

[8] [8]

Comparison of the substrate spectrum with fused silica The spectral shape of fused quartz (also called fused silica) from 300 to 600/cm is shown in Fig. S10. T he component Fig.2f , reproduced in Fig.S11 bottom left, appears to be a broadened version of this shape. Since the signal is very weak (1 pe maximum), the spectral shape of the component is less d...

2017