REVIEW 3 major objections 5 minor 9 references
Kalman filter/deep-learning hybrid automatic boundary tracking of optical coherence tomography data for deep anterior lamellar keratoplasty (DALK)
T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Adding a Kalman filter to deep-learning segmentation cuts average epithelium boundary error by 59.55% and Descemet's membrane error by 48.15% in ex vivo OCT M-mode tracking for DALK.
desk verdict A modest engineering application that plausibly reduces segmentation error on static OCT M-mode data, but the 'tracking' claim outruns the evidence without any temporal-response or lag analysis. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the Kalman filter integration: a one-dimensional state-space smoother that takes the U-Net's per-frame boundary detection as the observation $z_k$ and produces a smoothed boundary state $\hat{x}_{k|k}$. The prediction and update steps are $\hat{x}_{k|k-1}=F\hat{x}_{k-1|k-1}$ and $\hat{x}_{k|k}=\hat{x}_{k|k-1}+K_k(z_k-H\hat{x}_{k|k-1})$, with the paper's parameters $F=1$, $H=1$, $Q=10^{-5}$, $R=1$. That parameter choice lets the filter reject isolated jagged spikes while otherwise trusting the observations, and the adaptive sliding-window observation, 70% weight on the most recent 50 points and 30% on the preceding 50, turns the deep-learning boundary into a continuous, stable trace.
What would settle it
Run the proposed KDH pipeline on synthetic M-mode OCT sequences whose boundary moves as a known step or ramp, and compare the tracked depth against the known ground truth frame by frame. If the tracking error on fast-moving boundaries is much larger than the 0.28–0.36 pixel errors reported for the static ex vivo test set, the constant-position Kalman model is the reason.
Extended reading notes
Core claim
On its own terms, the paper's central discovery is that feeding a deep-learning network's boundary predictions into a Kalman filter with state transition $F=1$, observation $H=1$, process noise $Q=10^{-5}$, and observation noise $R=1$ substantially improves automatic tracking of the epithelium and Descemet's membrane in OCT M-mode scans from rabbit eyes undergoing Big Bubble DALK. The filter treats each boundary depth as a slowly time-varying random walk, uses the network prediction as a noisy observation, and outputs a smoothed depth trace. In the reported ex vivo comparison, the average epithelium error drops from 0.89 pixel (2.3229 µm) to 0.36 pixel (0.9396 µm), and the average DM error drops from 0.54 pixel (1.4094 µm) to 0.28 pixel (0.7308 µm), reductions of 59.55% and 48.15%, respectively. The authors also show qualitative examples where the filtered trace stays smooth under signal loss, motion artifacts, and jagged segmentation, and they demonstrate the pipeline in real time on an eye-mounted robotic OCT-guided needle system.
Load-bearing premise
The load-bearing assumption is that each boundary is almost stationary between frames, so the filter can treat any difference as noise; if the cornea moves quickly during needle insertion, the smoothed trace will lag behind the true boundary.
Editorial extensions
If this is right
- The average epithelium tracking error is reduced from 0.89 to 0.36 pixel (2.3229 µm to 0.9396 µm).
- The average DM tracking error is reduced from 0.54 to 0.28 pixel (1.4094 µm to 0.7308 µm).
- The filtered traces remain smooth and stable in low-signal, motion-artifact, and signal-loss cases where the deep-learning-only traces become jagged.
- The pipeline can be applied during real-time robotic-assisted DALK needle insertion, as demonstrated in the sample video.
- The resulting depth feedback is steadier than deep-learning-only segmentations, which is the property the authors argue is needed for surgical safety during Big Bubble DALK.
Reading between the lines
- A natural extension is to replace the constant-position random walk ($F=1$, small $Q$) with a constant-velocity or maneuver-aware model, since the paper does not analyze how much the filter would lag a fast-moving boundary during needle insertion.
- Because the reported error reductions are averaged over 50 test images, partitioning those images by signal quality would reveal whether the gain is entirely noise suppression; on high-SNR cases the filtered and unfiltered results already align, according to the paper's qualitative examples.
- The 70/30 sliding-window weighting is effectively a hand-tuned observation model; sweeping the $Q/R$ ratio and window weights on the existing test set would quantify how robust the 59.55% and 48.15% reductions are.
- The same Kalman wrapper is directly transferable to other OCT-guided needle procedures that track a single tissue boundary in real time, such as subretinal injection.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a Kalman filter/deep-learning hybrid (KDH) approach for tracking the Descemet membrane (DM) and epithelium boundaries in M-mode OCT images acquired during ex vivo DALK procedures. A U-Net produces boundary segmentations, and a scalar Kalman filter (F=1, H=1, Q=1e-5, R=1) with a sliding-window observation smoothing step is applied to the resulting boundary positions. On a test set of 50 patches from 12 rabbit eyes, the reported average absolute errors are 0.36 pixels (epithelium) and 0.28 pixels (DM) for KDH, versus 0.89 and 0.54 pixels for the deep-learning-only baseline, corresponding to 59.55% and 48.15% error reductions. The paper also provides qualitative examples and a video of real-time robotic-assisted needle insertion, and claims improved robustness in dynamic surgical conditions.
Significance. If the reported error reduction is robust and the dynamic tracking behavior is sound, the method could offer a simple, low-cost way to smooth deep-learning segmentation outputs in OCT-guided DALK, potentially improving boundary stability and surgical safety. The approach is easy to implement and builds directly on prior work by the same group. However, the quantitative evidence is limited to a single test set without uncertainty quantification, and the central claim of real-time tracking under motion is not validated by any temporal-response or lag analysis. The paper is an incremental but potentially useful engineering contribution; its practical significance depends on whether the Kalman smoother truly tracks boundary motion rather than merely smoothing static noise.
major comments (3)
- [Section 2.3 and Section 3.2] The Kalman configuration F=1, H=1, Q=1e-5, R=1 defines a constant-position random walk whose implied steady-state gain is K ≈ sqrt(Q/R) ≈ 3.2e-3, giving a response time constant on the order of 1/K ≈ 300 M-mode lines. The additional 70/30 sliding-window observation (introduced in Section 2.3) further low-pass filters the input. During needle insertion or tissue deformation, the DM and epithelium can move by several pixels over a few lines, and this filter configuration would lag such motion severely. The paper claims real-time tracking and robustness 'even under dynamic surgical conditions' (Section 3.2), but all quantitative results are on static M-mode images and no step-response, lag, or temporal-response analysis is provided. Consequently, the reported error reductions could be an artifact of smoothing jagged deep-learning outputs on quasi-static data, and the dynamic tracking claim is unsupported. Please add either (a) a temporal evaluation with known boundary displacement (e.g., simulated steps or annotated dynamic sequences during needle insertion) reporting lag and tracking error as a function of displacement speed, or (b) revise the claims to explicitly limit the method to smoothing of static boundary estimates.
- [Table 1 and Sections 3.1–3.2] The performance evaluation reports only point estimates of mean absolute error for a single test set of 50 patches from 12 rabbit eyes, with no error bars, confidence intervals, or significance tests. The claimed reductions of 59.55% and 48.15% are not accompanied by any measure of variability, so it is impossible to assess whether they are statistically reliable or how they vary across eyes or patches. The paper should report per-eye or per-patch error distributions (e.g., standard deviation, box plots), a paired significance test (e.g., Wilcoxon signed-rank or paired t-test), and a statement about whether the 50 test patches are independent (e.g., how many eyes contributed and whether multiple patches from the same eye are treated independently).
- [Section 3.2 and Abstract] The paper repeatedly claims 'real-time' capability for the KDH approach (e.g., abstract, Section 4) and states that the evaluation emphasized 'inference time efficiency,' yet no latency, throughput, or frame-rate numbers are reported anywhere. Without a measurement of the combined U-Net plus Kalman filter processing time on the stated GPU (NVIDIA GeForce RTX 4070 Ti SUPER) and a comparison to the OCT acquisition rate or the surgical update requirement, the real-time claim is not supported. Please include a timing table or, if the claim is intended only qualitatively, temper the language accordingly.
minor comments (5)
- [Section 1, paragraph 3] There is a duplicated and garbled passage: 'However, these methods are often challenged by signal noise, motion artifacts, and data inconsistencies, which can lead to jagged or inaccurate segmentation of the DM and epithelium boundary as shown in Figure 1 deep learning approaches like U-Net based approach , were employed to segment corneal layers in OCT M-scan data, enabling autonomous or semi-autonomous needle guidance. However, these methods are often challenged by signal' — this appears to be an editing error and should be rewritten as a single coherent sentence.
- [Figure 4 caption] The caption contains a placeholder URL 'http://dx.doi.org/doi.number.goes.here'; please replace it with a working link or remove the URL.
- [Section 2.3, equations] The Kalman equation notation is underspecified: please define x, z, K, and the covariance matrices explicitly, and state that all quantities are scalars in this application. In particular, explain how the sliding-window weighted average replaces the measurement z_k in Eq. (2) and whether the filter covariance P is updated in the usual way (the update equations for P are not shown).
- [Table 1] The column headers should indicate units explicitly (e.g., 'Average Epithelium Error (pixel/µm)') and state whether the errors are mean absolute errors averaged over all test patches, over M-mode lines, or per eye. This would improve reproducibility.
- [References] Reference [4] is a preprint (arXiv:2410.14577); if it has been published in a peer-reviewed venue, please update the citation. Also consider adding a reference for the Kalman filter itself in addition to the two application references [6,7] currently cited.
Circularity Check
No significant circularity: the Kalman-filter smoothing is evaluated against manual ground-truth labels, and the filter hyperparameters are set a priori rather than fitted to the test labels.
full rationale
The paper's derivation chain is empirical rather than definitional: the deep-learning segmentation produces boundary estimates, the Kalman filter smooths those estimates using fixed parameters (F=1, H=1, Q=1e-5, R=1, and a 70/30 sliding-window weighting), and the final boundaries are compared against manually annotated ground-truth labels from experienced ophthalmologists. The reported 59.55% and 48.15% error reductions are measured, not implied by the construction. The Kalman parameters and window weights are presented as chosen settings, not as quantities fitted to the test labels or derived from the ground truth, so the improvement is not forced by the method's own definitions. The paper cites several prior works by the same research group, but these citations are used for context, dataset provenance, and related surgical-system references; they do not supply a load-bearing theorem or a pre-fitted parameter that determines the reported results. No uniqueness theorem is invoked, no ansatz is smuggled in via citation, and no known result is merely renamed. The only substantive concern is a correctness risk, not circularity: the constant-position motion model with Q/R = 1e-5 may lag fast tissue motion during needle insertion, but that is a potential validity limitation, not a circular derivation. Therefore the circularity score is 1, reflecting routine self-citation without load-bearing circularity.
Assumptions & free parameters
free parameters (4)
- Process noise covariance Q =
1e-5
- Observation noise covariance R =
1
- Sliding window weights =
70% recent 50 points, 30% preceding 50 points
- Patch size =
16 x 512 x 32
assumptions (4)
- domain assumption Boundary depth evolves as a random walk with F=1, so the true DM and epithelium positions are roughly constant or slowly varying across M-mode frames.
- domain assumption The measurement noise on U-Net depth estimates is Gaussian with variance R=1.
- domain assumption Manual ophthalmologist annotations are correct ground truth for DM and epithelium boundaries.
- domain assumption Ex vivo rabbit eye OCT data are representative of human DALK surgical conditions.
Cite this review
Pith. "Pith review of Kalman filter/deep-learning hybrid automatic boundary tracking of optical coherence tomography data for deep anterior lamellar keratoplasty (DALK)." pith.science (2026). https://pith.science/paper/6CEIUHTY
@misc{pith2026250115311,
author = {Pith},
title = {Pith review of: Kalman filter/deep-learning hybrid automatic boundary tracking of optical coherence tomography data for deep anterior lamellar keratoplasty (DALK)},
year = {2026},
howpublished = {\url{https://pith.science/paper/6CEIUHTY}},
note = {Machine review of arXiv:2501.15311}
}
read the original abstract
Deep anterior lamellar keratoplasty (DALK) is a highly challenging partial thickness cornea transplant surgery that replaces the anterior cornea above Descemet's membrane (DM) with a donor cornea. In our previous work, we proposed the design of an optical coherence tomography (OCT) sensor integrated needle to acquire real-time M-mode images to provide depth feedback during OCT-guided needle insertion during Big Bubble DALK procedures. Machine learning and deep learning techniques were applied to M-mode images to automatically identify the DM in OCT M-scan data. However, such segmentation methods often produce inconsistent or jagged segmentation of the DM which reduces the model accuracy. Here we present a Kalman filter based OCT M-scan boundary tracking algorithm in addition to AI-based precise needle guidance to improve automatic DM segmentation for OCT-guided DALK procedures. By using the Kalman filter, the proposed method generates a smoother layer segmentation result from OCT M-mode images for more accurate tracking of the DM layer and epithelium. Initial ex vivo testing demonstrates that the proposed approach significantly increases the segmentation accuracy compared to conventional methods without the Kalman filter. Our proposed model can provide more consistent and precise depth sensing results, which has great potential to improve surgical safety and ultimately contributes to better patient outcomes.
Reference graph
Works this paper leans on
-
[1]
Gensheimer, W. G., Opfermann, J., Wang, Y., Kaluna, J., Krieger, A., and Kang, J. U., “Comparison of vertical cannula insertion techniques for big bubble deep anterior lamellar keratoplasty,” Investigative Ophthalmology & Visual Science 65(7), 3700–3700 (2024)
work page 2024
-
[2]
Singh, M. S., Li, K., Yu, J., Liu, X., Kang, J. U., and others, “Live porcine eye model studies of subretinal injection using handheld endoscopy OCT integrated injector,” Investigative Ophthalmology & Visual Science 65(7), 5499 –5499 (2024)
work page 2024
-
[3]
Opfermann, J., Wang, Y., Kaluna, J., Krieger, A., Kang, J. U., and Gensheimer, W. G., “Novel vertical big bubble needle approach vs. the standard approach for deep anterior lamellar keratoplasty (DALK),” Investigative Ophthalmology & Visual Science 65(7), 3684–3684 (2024)
work page 2024
-
[4]
Wang, Y., Opfermann, J. D., Yu, J., Yi, H., Kaluna, J., Biswas, R., Zuo, R., Gensheimer, W., Krieger, A., and Kang, J. U., “Reimagining partial thickness keratoplasty: An eye mountable robot for autonomous big bubble needle insertion,” arXiv preprint arXiv:2410.14577 (2024)
arXiv 2024
-
[5]
Common -path optical coherence tomography guided vertical pneumodissection for DALK,
Wang, Y., Guo, S., Opfermann, J. D., Kaluna, J., Gensheimer, W. G., Krieger, A., and Kang, J. U., “Common -path optical coherence tomography guided vertical pneumodissection for DALK,” Optical Fibers and Sensors for Medical Diagnostics, Treatment and Environmental Applications XXIII 12372, 15 –19 (2023)
work page 2023
-
[6]
Dual -kernel driven convolutional sparse learning for bearing transient feature enhancement,
Yi, H., Ding, X., Li, Q., Wang, H., Tang, J., Liu, R., and Huang, W., “Dual -kernel driven convolutional sparse learning for bearing transient feature enhancement,” Measurement 216, 112643 (2023)
work page 2023
-
[7]
Integrated treatment planning in percutaneous microwave ablation of lung tumors,
Wang, H., Yi, H., Liu, J., and Gu, L., “Integrated treatment planning in percutaneous microwave ablation of lung tumors,” 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 4974–4977 (2022)
work page 2022
-
[8]
Yu, J., Yi, H., Wang, Y., Opfermann, J. D., Gensheimer, W. G., Krieger, A., and Kang, J. U., “Topology -based deep-learning segmentation method for deep anterior lamellar keratoplasty (DALK) surgical guidance using M -mode OCT data,” arXiv preprint arXiv:2501.04735 (2025)
work page Pith review arXiv 2025
Show all 9 references
-
[9]
The neural networks -based needle detection for medical retinal surgery,
Xu, J., Yu, J., Yao, J., and Zhang, R., “The neural networks -based needle detection for medical retinal surgery,” International Conference on Computer Graphics, Artificial Intelligence, and Data Processing (ICCAID 2022) 12604, 674–678 (2023)
2023
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.