{"id":"e3611b7e-f837-481f-b6e5-38f91f1317c9","arxiv_id":"2501.01752","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A regression network can locate the sensing area of a laparoscopic gamma probe on tissue from RGB images and probe-axis points, trained with a laser-annotated mock probe.","lead":"This PhD thesis develops computer vision tools to show surgeons where a handheld gamma probe is detecting radiation during keyhole cancer surgery. It tests a regression network that locates the probe's sensing area on tissue from standard laparoscopic images, using a laser spot as training guidance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 70.5-pixel error may reflect interpolation between phantom rotations rather than generalization to unseen camera-probe poses: the dataset split in §7.4.4 appears random at image level, not pose-disjoint, so each pose appears in both training and test.","rationale":"The reader's weakest assumption concerned the laser mock-probe proxy and generalization to real tissue/geometry. I agree that this is a genuine limitation, but the more immediate and testable weakness is the evaluation protocol: without a pose-disjoint split, the reported numbers do not establish that the network generalizes to novel camera-probe poses, which is precisely what intraoperative use requires. This is not an accusation of misconduct; it is a missing experimental control. The proposed concrete test would settle the matter. Since the reader's verdict was already CONDITIONAL and this concern reinforces that assessment, no verdict change is needed.","tokens_in":48742,"tokens_out":4799,"duration_ms":49095,"concrete_test":"Re-run the best ResNet+MLP configuration from Table 7.2 with a pose-disjoint split: train on 100 of the 120 camera-probe poses, validate on 10, test on the remaining 10 held-out poses (and likewise for Coffbee), and report mean/median 2D error and R². If held-out-pose error is substantially higher than 70.5 px (e.g., >100 px or R²<0.6), the benchmark claim is unsupported; additionally, test on a phantom surface excluded from training to rule out texture memorization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Chapter 7 is that a simple regression network can infer the SENSEI probe's tissue intersection point and 'establish a new benchmark'. The most load-bearing condition for that claim is that the reported errors measure generalization to new intraoperative views. The dataset construction (§7.3) yields 120 camera-probe poses × 10 phantom surface profiles = 1200 images (Jerry; similarly Coffbee). Section 7.4.4 states the data were 'partitioned' into 800/200/200 train/validation/test images. If this partition is at image level—and the text gives no indication of stratification by pose—then each camera-probe pose appears in both training and test. The network can exploit pose-specific image cues (probe location, orientation, phantom texture) and interpolate across the 10 phantom rotations of the same pose, so the 70.5 px mean error and R²=0.82 in Table 7.2 overstate performance on unseen poses. The abstract's 'new benchmark' is therefore not yet supported. A secondary issue is that the laser spot is used as ground truth for the gamma probe's sensing area without reporting calibration of laser-probe axis alignment; but the pose-split problem is more fundamental because it affects the validity of every reported number.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This PhD thesis addresses the intraoperative visualisation of a tethered laparoscopic gamma probe (SENSEI) in minimally invasive cancer surgery. It develops marker-based probe tracking with augmented reality (Chapter 3), self-supervised depth estimation methods SADepth and M3Depth (Chapters 4 and 5), joint depth estimation and surgical tool segmentation (Chapter 6), and a new sensing-area detection method based on a laser-instrumented mock probe and a regression network (Chapter 7). The central claim of Chapter 7 is that a simple ResNet-plus-MLP network can predict the probe-tissue intersection point in real time from standard RGB stereo images together with PCA-sampled probe-axis points, reporting a best 2D mean error of 70.5 pixels, a 3D mean error of 7.4 mm, an R2 score of 0.82, and an inference rate of 50 frames per second, and describing this as a new benchmark for the surgical vision community.","tokens_in":49034,"tokens_out":5678,"duration_ms":53764,"significance":"The thesis contains three previously published method chapters and a new application chapter. Its strongest assets are the problem reformulation that turns the sensing-area localisation task into a 2D regression problem, the custom hardware platform with structured-light depth references, and the public release of the Jerry and Coffbee datasets. If the reported accuracy held under a pose-disjoint evaluation, the simple ResNet-plus-MLP architecture at 50 fps would be a practically useful contribution to intraoperative visualisation. However, the evaluation as presented does not yet establish generalisation to unseen camera-probe poses, and the claimed 'new benchmark' is therefore premature.","major_comments":[{"comment":"The evaluation protocol does not support the generalisation claim. Section 7.3 states that each of the 120 camera-probe poses is imaged under 10 phantom surface profiles, yielding 1200 images, and Section 7.4.4 partitions the Jerry and Coffbee datasets into 800/200/200 images without any statement of stratification by pose or surface profile. With an image-level random split, frames from the same pose can appear in both training and test, so the reported 70.5 px mean error and R2=0.82 may reflect interpolation across the 10 rotations of a seen pose rather than detection on unseen poses. Please report results with a pose-disjoint split, for example training on 100 poses and testing on the remaining 20 poses, and state whether the 70.5 px figure changes.","section":"§7.3, §7.4.4"},{"comment":"The claimed 'new benchmark' is not supported without a baseline comparison to the geometric tracking and SfM intersection method of Chapter 3. Section 7.4.1 rejects that approach on practical grounds such as sterilisation and the need to move the laparoscope, but it provides no quantitative comparison. Since Chapter 3 already reports pose-estimation errors and an AR intersection demonstration, the sensing-area task can be evaluated against that geometric approach using the same error metric. Please add this comparison, and ideally also the segmentation baselines of Table 7.1 under the same test protocol.","section":"§7.4.1, §7.5, Table 7.2"},{"comment":"The laser module is used to define ground truth for the gamma probe sensing area, but the manuscript reports no calibration of the alignment between the laser beam axis and the SENSEI probe axis, and no sensitivity analysis of how laser-spot-centre errors propagate to the regression labels. Because the laser spot centroid is the training target, any laser-to-axis misalignment in the mock probe or in a real probe corrupts both training and test labels. Please report the calibration procedure and, ideally, a tolerance analysis.","section":"§7.3, §7.4.3"},{"comment":"The text states that 'LSTM and MLP gave competitive results and they are all in sub-millimetre level,' but Table 7.2 lists 3D mean errors of 6.4-11.2 mm and medians of 4.0-5.4 mm. This is an internal inconsistency that should be corrected. The 3D error metric should also be described more carefully given the depth-map resolution and missing-depth limitations noted in the same section.","section":"§7.5, Table 7.2"}],"minor_comments":[{"comment":"The R2 score and the 3D error rows are not explicitly associated with the grey (Jerry) and blue (Coffbee) dataset blocks; please clarify the table layout in the caption.","section":"Tables 7.2 and 7.3"},{"comment":"The statement that the method 'relies solely on the 2D information and works well without the need for the laser module after training' is not accompanied by a failure-mode analysis; consider reporting cases where PCA-based axis extraction fails, such as low-texture tissue or partial probe occlusion.","section":"§7.4.1"},{"comment":"The notation in Eq. (5.7) is inconsistent: the subscripts and superscripts of L^{2D(l)}_{lr}, L^{2D(r)}_{lr}, and L^{3D}_{gc} are not defined uniformly, and L^{3D}_{lr} appears without a definition.","section":"Eq. (5.7)"},{"comment":"The text says 'Code and data are available at this link,' but no URL or repository identifier appears in the manuscript; please provide the actual link.","section":"Chapter 7"}],"recommendation":"major_revision","confidential_remarks":"The pose-split issue is the main risk: the central claim of a benchmark cannot be evaluated until a pose-disjoint test is reported. The thesis is largely a compilation of the author's previously published papers, so the editor may wish to consider whether the novel contribution of Chapter 7 is sufficiently validated for a journal submission in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea in Chapter 7 is real: casting gamma probe sensing area detection as regression from stereo RGB plus sampled probe-axis points, trained with laser-derived labels, is not something I've seen elsewhere. The hardware setup (stereo laparoscope, rotation stage, laser mock probe, shutter control) is thoughtful, and the two datasets, if actually released, could be genuinely useful to the surgical vision community. The earlier depth-estimation chapters are competent Monodepth-style extensions, already peer-reviewed at MICCAI and T-MRB; they're not the main event but they hold up.\n\nThe soft spot is the validation of the central claim. Section 7.4.4 says the 1200 images (120 camera-probe poses x 10 phantom surface rotations) are partitioned into 800/200/200 without any mention of stratifying by pose. If that split is random at image level, then each pose appears in both training and test, and the reported 70.5-pixel mean error / R2=0.82 mostly measure interpolation across the 10 surface rotations of a fixed pose. That is not the same as generalizing to unseen camera-probe poses, which is what the 'new benchmark' language implies. The stress-test note is right to flag this. I'm not claiming the method is bad; I'm claiming the reported numbers don't yet mean what the abstract says they mean.\n\nThere is also no quantitative comparison to the geometric tracking and SfM baseline described in Chapter 3. That is the natural null method to beat, and its absence is conspicuous. The validation is phantom-only, with a mock laser probe; the alignment between laser spot and the real gamma probe's sensing axis is never calibrated. Those limitations are not fatal for a conference-grade contribution, but they further cap what the current evidence supports.\n\nMinor stuff: the code/data link is promised but not actually present in the arXiv rendering, and at least two reference labels look wrong. Both are easy fixes.\n\nWho gets value from this: researchers working on surgical instrument guidance and intraoperative visualization, especially anyone who wants to pick up the Jerry/Coffbee datasets for the new task. The problem formulation is worth having in the literature.\n\nRecommendation: don't desk reject. Send it to peer review with explicit reviewer instructions to demand a pose-disjoint evaluation and a comparison against the geometric baseline. If those are added, the benchmark claim becomes testable; without them, the abstract's 'new benchmark' is unsupported.","headline":"Novel problem formulation and a genuinely useful dataset, but the benchmark claim rests on a train/test split that likely lets the network interpolate between phantom rotations of the same pose rather than generalize to new poses.","tokens_in":49541,"tokens_out":2572,"would_cite":false,"duration_ms":29448,"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":"This thesis shows a simple regression network can predict a laparoscopic gamma probe's sensing spot on tissue directly from standard RGB images and probe-axis points.","keywords":["gamma probe","sensing area detection","laparoscopic surgery","surgical vision","regression network","radio-guided surgery","probe axis-surface intersection","intraoperative visualisation"],"falsifier":"Mount a real gamma probe with a fixed laser module so both axes are aligned, place a small radioactive source under a phantom, scan the probe across the surface, and record where the gamma count peaks; compare that peak location with the network's predicted intersection point and with the laser spot. If the predicted point is consistently off by more than the reported 3D error while the laser spot agrees with the gamma peak, the method's labels are wrong; if the laser spot itself disagrees with the gamma peak, the mock-probe assumption collapses.","tokens_in":48503,"feed_emoji":"🎯","tokens_out":7978,"duration_ms":78127,"temperature":0.7,"pith_summary":"This thesis argues that a surgeon using a tethered laparoscopic gamma probe during minimally invasive cancer surgery can be shown, in real time, exactly which spot on the tissue surface the probe is sensing. The central move is to stop treating the task as a 3D geometry problem — reconstructing the tissue surface and the probe pose — and instead treat it as a 2D inference problem: predict the probe-axis/tissue intersection point directly from an ordinary laparoscopic RGB image plus a few points sampled along the probe's apparent axis. To get training labels, the author replaces the probe's sensing tip with a laser-module mock probe, so the visible laser spot marks the ground-truth intersection; after training, the laser is no longer needed. The best configuration, a ResNet image branch combined with a simple MLP for the axis points, runs at 50 frames per second and achieves the reported benchmark on two newly released datasets. If this holds in the operating theatre, it would give surgeons a direct visual cue of where gamma signals originate, potentially reducing incomplete resections and enabling augmented-reality guidance.","feed_headline":"A simple network reveals where a gamma probe is sensing","feed_subtitle":"Surgeons could see the probe-tissue intersection live, replacing guesswork from audible gamma counts.","key_machinery":"The load-bearing object is the laser-module mock probe: a non-functional gamma-probe shell fitted with a red laser whose visible spot on the phantom marks the probe axis-tissue intersection, providing the ground truth label. The load-bearing method is a two-branch regression network: a ResNet50 branch encodes the laser-off stereo RGB image, an MLP branch encodes 50 points sampled along the probe axis extracted by PCA, the two feature vectors are concatenated, and an MSE loss regresses the 2D intersection point. Paired laser-on/laser-off image acquisition, an electrically controlled shutter, and a rotation stage generate the Jerry and Coffbee datasets, with structured-light depth maps in Coffbee supplying 3D error evaluation. The identity that carries the argument is that the laser spot centre equals the sensing-area ground truth, so training on laser-off images transfers to the real probe situation where no spot exists.","core_discovery":"The paper's central claim is that the sensing area of a laparoscopic gamma probe — defined as the intersection of the probe axis with the tissue surface, projected into the 2D laparoscopic image — can be predicted directly from the standard 2D image and the probe's apparent axis, without per-pixel depth, without tracking markers, and without the laser module at inference time. The problem is reformulated as laser-point inference: a DAQ-controlled laser module mounted in a non-functional gamma-probe shell emits a visible spot that marks the intersection, paired laser-on/laser-off stereo images supply ground truth labels, and a two-branch regression network (ResNet for image features, MLP for 50 PCA-sampled principal-axis points) is trained end-to-end with mean squared error loss on laser-off images. Segmentation baselines fail because they depend on the laser spot being present; the regression approach succeeds because the network learns scene-level cues from the image and probe geometry. On the two newly acquired datasets the ResNet+MLP combination reports a mean 2D error around 70 pixels, an $R^2$ of 0.82, and 50 fps inference, which the thesis describes as a new benchmark for the surgical vision community.","pith_inferences":["Extension: a decisive test not performed in the thesis would use a real gamma probe with a buried radioactive source and compare the predicted sensing point with the point of maximum gamma count; if the laser axis and gamma detector axis are not co-aligned in the clinical device, the reported accuracy will not transfer.","Extension: the phantom-to-patient gap remains the main open question, since silicone tissue with hand-painted colour has more texture than many in-vivo scenes and PCA-based axis extraction may degrade on low-texture, specular, or blood-covered tissue; evaluating on in-vivo or ex-vivo video with laser labels would bound the domain shift.","Extension: because the network takes the 2D probe axis as input, a tracking failure that slightly rotates the axis estimate will propagate directly into the intersection prediction; fusing kinematic data from the robotic arm could stabilise this in practice, an option the thesis mentions only as future work."],"forward_implications":["If the central claim holds, intraoperative AR overlays showing the gamma probe's sensing area can be generated from a standard RGB laparoscope feed at 50 fps, with no extra tracking hardware.","Surgeons would no longer need to memorise count readings; a visible cue on tissue could reduce incomplete resections and unnecessary dissection of healthy lymph nodes.","The paired laser-on/laser-off protocol provides a practical way to generate intersection-point ground truth for other non-imaging probes, such as ultrasound or diffuse reflectance spectroscopy probes.","The two released datasets give other groups a common benchmark to compare sensing-area detection algorithms."],"supporting_citations":[{"why":"Supplies the ResNet50 image-feature backbone used in the best-performing regression configuration.","marker":"[144]"},{"why":"Provides the Vision Transformer baseline that the comparison shows performs worse than ResNet for this task.","marker":"[196]"},{"why":"Provides the LSTM alternative for encoding the sampled probe-axis points, compared against the simpler MLP.","marker":"[197]"},{"why":"Swin-UNet segmentation baseline used in the first failed attempt; its predictions are invalid on a large share of frames.","marker":"[193]"},{"why":"TransUNet segmentation baseline that fails when the laser spot is absent, motivating the regression formulation.","marker":"[194]"},{"why":"PCA is used to extract the probe's central axis from the image and to sample the 50 principal points fed to the MLP branch.","marker":"[195]"},{"why":"Provides the structured-lighting depth ground truth used to build the Coffbee dataset and evaluate 3D errors.","marker":"[3]"}],"fun_headline_variants":["AI predicts gamma probe sensing point on tissue in real time","Neural net shows where gamma probe touches tissue live","Probe-tissue intersection predicted from 2D video alone","Regression beats segmentation for gamma probe localization","Deep learning maps gamma probe field to tissue surface"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The mock probe's laser beam marks exactly the same tissue point that the real gamma probe would sense, and a network trained on 120 silicone-phantom poses with that laser label will transfer to real tissue and real probe geometry.","fun_headline_variants_meta":{"raw":{"variants":["AI predicts gamma probe sensing point on tissue in real time","Neural net shows where gamma probe touches tissue live","Probe-tissue intersection predicted from 2D video alone","Regression beats segmentation for gamma probe localization","Deep learning maps gamma probe field to tissue surface"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3095,"prompt_tokens":1060,"completion_tokens":2035,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":1970}},"tokens_in":676,"tokens_out":2035,"duration_ms":12589,"temperature":1.0,"reasoning_tokens":1970,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:20:51.379240+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Mount a real gamma probe with a fixed laser module so both axes are aligned, place a small radioactive source under a phantom, scan the probe across the surface, and record where the gamma count peaks; compare that peak location with the network's predicted intersection point and with the laser spot. If the predicted point is consistently off by more than the reported 3D error while the laser spot agrees with the gamma peak, the method's labels are wrong; if the laser spot itself disagrees with the gamma peak, the mock-probe assumption collapses.","supporting_citations":[{"cited_title":"Swin-unet: Unet-like pure transformer for medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Swin-UNet segmentation baseline used in the first failed attempt; its predictions are invalid on a large share of frames."}],"review_version":1}