{"id":"0c098db3-927d-40bd-bd21-1d3108b31294","arxiv_id":"2501.04735","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A star-shaped topology loss plus a modified U-Net improves M-mode OCT segmentation of corneal layers, though the evaluation is limited by subject-wise data leakage and missing error bars.","lead":"Researchers tested a topology-based loss function with a modified U-Net for segmenting corneal layers in M-mode OCT images used in robot-assisted cornea surgery. The method reports faster and more accurate layer tracking than the team's previous approach, but the evaluation has data-splitting limits and no statistical testing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random train/test split without grouping by individual eye lets the network memorize eye-specific patterns, so the reported accuracy gains may not transfer to unseen eyes; the authors' own Discussion concedes this.","rationale":"The reader identified the random split without eye grouping as the weakest assumption, and the manuscript's own Discussion confirms it. This is the single most load-bearing issue because the clinical claim is about guiding surgery on new eyes; if test frames come from the same eyes seen in training, the reported accuracy is an upper bound on real-world performance. The proposed check directly tests this by forcing evaluation on entirely unseen eyes. I found no internal inconsistency in the loss or architecture that would be more fundamental, and the paper does provide reproducible-sounding experimental comparisons, though without code or significance tests. The verdict should remain conditional pending the unseen-eye evaluation.","tokens_in":5388,"tokens_out":2769,"duration_ms":25038,"concrete_test":"Regroup all images by individual eye (the in vivo set has 4 rabbit eye pairs, the ex vivo set 12 rabbit eyes) and retrain the proposed method and the conventional baseline under the same hyperparameters using leave-one-eye-out cross-validation. Compare average SSIM, IoU, Dice, and boundary errors on the held-out eyes against Tables 1 and 2. If the proposed method's advantage largely disappears or reverses on unseen eyes, the central claim of robust new-eye guidance is not supported; if the advantage persists, the random-split concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the proposed method provides robust, real-time guidance for DALK on corneal layers, which requires the model to segment eyes it has not trained on. The evaluation instead randomly splits 512x512 M-mode images into training and test sets (Sec. 3.1) without grouping by individual eyeball. Each eye contributes many time-series frames, so frames from the same eye appear on both sides of the split. The network can learn eye-specific artifacts, noise profiles, and layer geometry, making the reported SSIM ~0.99 and sub-voxel tracking errors (Tables 1-2) reflect within-eye memorization rather than cross-eye generalization. The authors explicitly acknowledge this in the Discussion: 'the current data set is randomly split, allowing the network to potentially learn from similar samples of the same eyeball, it is essential to evaluate the method on entirely unseen eyeball data in next stage experiments.' This is a limitation of the evidence rather than an internal inconsistency, but it is the weakest load-bearing link for the guidance claim. Secondary issues (no error bars, only one baseline) reinforce the need for a held-out-eye evaluation but do not by themselves invalidate the reported comparisons.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning segmentation method for M-mode OCT images of the cornea, intended to guide deep anterior lamellar keratoplasty (DALK). The method combines a binary cross-entropy loss with a star-shape-prior topological loss and a modified U-Net architecture. The authors evaluate the method on in vivo, ex vivo, and hybrid rabbit eye data sets, reporting improvements over their previous U-Net with BCE loss in SSIM, PSNR, IoU, Dice, and inference frequency, plus reduced average absolute tracking errors for the epithelium and Descemet's membrane. The paper closes by acknowledging that the random data split may let the network learn from similar samples of the same eyeball.","tokens_in":5589,"tokens_out":2463,"duration_ms":26390,"significance":"If the reported improvements survive a proper held-out-eye evaluation, the method is a useful incremental contribution to real-time OCT-guided surgical robotics: the star-shape topological loss is a principled way to enforce geometric continuity under noise, and the inference speed gain (35–40 Hz vs 25–28 Hz) is practically relevant. The main value is in demonstrating that a topology-aware loss can stabilize layer segmentation in M-mode OCT, a domain where conventional pixel-wise losses produce fragmented boundaries. The principal weakness is that the current evaluation does not demonstrate generalization to unseen eyes, which is the actual clinical requirement. The ablation structure also does not isolate the contribution of the topological loss from the architectural changes.","major_comments":[{"comment":"The train/test split is performed at the level of 512×512 M-mode frames without grouping by individual eyeball, so frames from the same eye appear in both the training and test sets. Since M-mode frames from one eye are highly correlated, the strong metrics in Tables 1 and 2 likely reflect within-eye memorization rather than cross-eye generalization. The authors explicitly concede this in the Discussion: 'the current data set is randomly split, allowing the network to potentially learn from similar samples of the same eyeball.' Because the central claim is robust real-time guidance for DALK, a held-out-eye split (e.g., training on some eyes and testing on entirely different eyes) is a required experiment before the claim can be accepted.","section":"Section 3.1"},{"comment":"The only comparison is against the authors' previous U-Net architecture with BCE loss. Since the proposed method changes both the network architecture and the loss function, the reported improvements cannot be attributed to the topological loss alone. An ablation study is needed: for example, the modified U-Net trained with BCE only, or the previous U-Net trained with the hybrid loss, would isolate the effect of the loss function and the architecture. Without such an ablation, the central claim that the topology-based loss is the source of robustness is not supported.","section":"Section 3.2"},{"comment":"No error bars, confidence intervals, or statistical significance tests are reported for any metric. The test sets are small (100, 50, and 150 images for in vivo, ex vivo, and hybrid, respectively), and the reported averages could hide substantial run-to-run or sample-to-sample variation. For example, the ex vivo PM and CM IoU values (0.9874 vs 0.9798) differ by less than 0.01, and without variance information it is unclear whether this difference is reproducible. Reporting standard deviations or performing paired significance tests across test images would strengthen the quantitative claims.","section":"Section 3.2, Tables 1 and 2"}],"minor_comments":[{"comment":"The topological loss formula uses the sets O and the center c without explicitly defining them in the text; a precise definition of the star-shape region and how c is determined would improve reproducibility.","section":"Section 2.2, Eq. (3)"},{"comment":"The paper states α/β = 1/2 but does not specify the actual values of α and β or describe how they were chosen; adding these details would clarify the experimental setup.","section":"Section 2.2"},{"comment":"The statement that the modified U-Net 'adopts a more stable and precise design' is qualitative; quantitative details such as the number of feature channels, depth, and the exact placement of transposed convolutions and normalization are omitted.","section":"Section 2.3"},{"comment":"The video link in Fig. 5 is a placeholder ('http://dx.doi.org/doi.number.goes.here') and should be replaced with the actual DOI or URL.","section":"Section 3.2"},{"comment":"Reference [1] is cited as the source of the previous U-Net architecture, but the title of [1] ('Comparison of vertical cannula insertion techniques for big bubble deep anterior lamellar keratoplasty') suggests it is a clinical abstract, not an architecture paper; a more appropriate citation for the baseline network should be provided.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an honest empirical study with a clear limitation that the authors themselves acknowledge. The main technical risk is not internal inconsistency but evaluation validity: the random split by frame rather than by eye is the weakest link for the clinical guidance claim. The comparison to only the authors' own prior baseline, without ablations, further limits the attribution of the improvement to the topological loss. Both issues are fixable with additional experiments, and the paper's scope is appropriate for a specialized medical-imaging venue. There is also a moderate concern that several references are self-citations to conference abstracts that may not provide full methodological detail for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a competent engineering paper with an honest limitation that the authors themselves name. The new thing is not the loss — that's Mirikharaji and Hamarneh's star shape prior from 2018 — but the successful transfer to M-mode OCT corneal layer segmentation for DALK guidance. The modified U-Net is incremental. The reported numbers are consistent and the qualitative examples show real improvement in noisy M-mode scans. The method reaches 35-40 Hz inference, which matters for the robot guidance use case. I'd give credit for doing the application and for being clear about the evaluation gap.\n\nThe main soft spot is the data split. Frames from the same eye appear in both training and test, so the model can memorize eye-specific texture and noise. The authors acknowledge this in the Discussion: 'allowing the network to potentially learn from similar samples of the same eyeball.' That is a load-bearing weakness for the central claim of robust real-time surgical guidance on unseen eyes. The SSIM ~0.99 and sub-pixel tracking errors are probably optimistic. Also no error bars or significance tests, and the baseline is only their own prior network. Those are secondary but reinforcing.\n\nI don't see a circularity problem. The topological loss is independent of the baseline. The alpha/beta weights are fixed; no fitted parameters are hidden in the claim. The placeholder URL for the video is a minor editorial issue.\n\nWho gets value: anyone working on OCT-guided ophthalmic surgery or learning-based segmentation with shape priors. It's a useful data point that star-shape loss helps in noisy M-mode data. Not a foundational result, but a legitimate incremental contribution.\n\nRecommendation: this deserves peer review — a serious referee could test the held-out-eye question quickly — but I'd push for a re-evaluation on eye-disjoint splits and at least error bars before accepting.","headline":"A plausible incremental application of a star-shaped topological loss to M-mode OCT corneal segmentation, but the random train/test split undermines the generalization claim.","tokens_in":6125,"tokens_out":1982,"would_cite":false,"duration_ms":18715,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that adding a star-shape topological loss to a modified U-Net makes M-mode OCT segmentation of the epithelium and Descemet's membrane accurate and fast enough for real-time DALK guidance.","keywords":["deep anterior lamellar keratoplasty","DALK","Descemet membrane","optical coherence tomography","M-mode OCT","image segmentation","topological loss","star shape prior"],"falsifier":"A leave-one-eye-out test: train on all images from some rabbit eyes, test only on images from a held-out eye, and compare the average Descemet's membrane boundary error; if it exceeds the reported 1.7 µm in vivo by a large margin, the robustness claim as stated for unseen eyes fails.","tokens_in":5188,"feed_emoji":"👁","tokens_out":7742,"duration_ms":69895,"temperature":0.7,"pith_summary":"The paper is trying to establish that a deep-learning segmentation method built around a topological loss can reliably track two corneal layers—the epithelium and Descemet's membrane—from noisy M-mode OCT signals during robotic DALK surgery. The authors combine a pixel-wise cross-entropy term with a star-shape-prior topological term and a faster modified U-Net, and they report that this hybrid beats the conventional BCE-loss U-Net on every metric and on every data set tested. The motivation is surgical: the Big Bubble technique in DALK requires micron-accurate separation of deep stroma from Descemet's membrane, and real-time guidance needs both accuracy and speed. If the claim holds, the method would make the OCT-guided robot's layer tracking stable enough for clinical use.","feed_headline":"Topology-aware network tracks corneal layers at 40 Hz","feed_subtitle":"A star-shape loss keeps corneal layer boundaries accurate under noisy M-mode OCT for DALK guidance.","key_machinery":"Two pieces carry the argument. The first is the hybrid loss $L_{\\mathrm{hybrid}} = \\alpha L_{\\mathrm{BCE}} + \\beta L_T$ with $\\alpha/\\beta=1/2$: $L_T$ encodes a star shape prior, requiring that along the straight line from any pixel inside the object to the region center, all pixels also belong to the object, which penalizes the holes and rough boundaries that noise produces. The second is a modified U-Net using smaller 3×3 convolution kernels, learnable upsampling with transposed convolutions, and normalization; this speeds up inference without losing spatial resolution. The paper also crops each 512×512 image into 8×512×64 patches for training and reconstructs the patches into the original format during inference.","core_discovery":"The central claim is that incorporating a star-shape-prior topological loss into a modified U-Net makes corneal layer segmentation from M-mode OCT robust to the noise, signal instability, and motion artifacts that arise when an OCT-fiber needle moves inside the eye. The authors state this as a concrete result: on in vivo, ex vivo, and hybrid rabbit-eye data sets, the proposed method achieved higher SSIM, PSNR, IoU, and Dice coefficients and higher inference frequency (up to 40 Hz) than the conventional method, while reducing average absolute boundary tracking errors—down to 0.54 pixels (1.41 µm) for the epithelium and 0.64 pixels (1.67 µm) for Descemet's membrane in vivo.","pith_inferences":["The paper does not separate the contributions of the loss and the architecture; an ablation that removes the topological term or reverts to the old U-Net would show which part drives the improvement.","Because the evaluation splits images rather than eyeballs, the reported numbers are an upper bound on cross-eye generalization; a leave-one-eye-out evaluation would be a direct test of the generalization claim.","The star-shape prior appears naturally suited to other layered or convex anatomical targets in M-mode OCT, such as retinal layers or vessel walls, where the same loss could be reused with minimal changes.","If the speed and accuracy numbers hold during needle motion, the robot could plausibly close a control loop on the tracked Descemet's membrane boundary rather than relying on a pre-planned insertion depth."],"forward_implications":["The reported inference frequency reaches 35–40 Hz, fast enough to drive real-time robot control during surgery.","Boundary tracking errors drop to roughly half a pixel in vivo, about 1.4–1.7 µm, which matches the micron-scale precision the Big Bubble technique needs.","The same hybrid-loss design should transfer to other noisy OCT-guided procedures where layer continuity matters more than pixel-level purity.","The loss term does not change the annotation protocol, so existing labeled M-mode OCT data can be reused to retrain the network."],"supporting_citations":[{"why":"Defines the star-shape-prior topological loss that forms the core of the hybrid objective.","marker":"[9]"},{"why":"Supplies the in vivo M-mode OCT data from the prior robot-assisted DALK experiment and the earlier segmentation approach treated as the conventional baseline.","marker":"[8]"},{"why":"Describes the eye-mountable robotic injection system whose real-time layer tracking motivates the method and provides the demonstration video.","marker":"[7]"}],"fun_headline_variants":["Topology loss makes OCT corneal tracking noise-proof","40 Hz corneal boundary tracking with topology-aware AI","Star-shape loss sharpens corneal layer segmentation","Noise-proof OCT corneal layer detection for DALK","Topology-aware AI stabilizes DALK needle guidance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation randomly splits M-mode images into training and test sets without keeping each rabbit eyeball entirely in one side, and the authors acknowledge this can let the network learn from similar frames of the same eye, raising the measured performance beyond what a new eye would give.","fun_headline_variants_meta":{"raw":{"variants":["Topology loss makes OCT corneal tracking noise-proof","40 Hz corneal boundary tracking with topology-aware AI","Star-shape loss sharpens corneal layer segmentation","Noise-proof OCT corneal layer detection for DALK","Topology-aware AI stabilizes DALK needle guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3139,"prompt_tokens":907,"completion_tokens":2232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2159}},"tokens_in":523,"tokens_out":2232,"duration_ms":16153,"temperature":1.0,"reasoning_tokens":2159,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:39:50.094460+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A leave-one-eye-out test: train on all images from some rabbit eyes, test only on images from a held-out eye, and compare the average Descemet's membrane boundary error; if it exceeds the reported 1.7 µm in vivo by a large margin, the robustness claim as stated for unseen eyes fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the star-shape-prior topological loss that forms the core of the hybrid objective."},{"cited_title":"D., Wang, Y., Kang, J","cited_arxiv_id":null,"evidence_quote":"Describes the eye-mountable robotic injection system whose real-time layer tracking motivates the method and provides the demonstration video."}],"review_version":1}