{"id":"29ca9890-70d5-4e71-868c-cafde0757ba0","arxiv_id":"1908.01543","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An automatic CT pipeline combining a spatially aware neural network, tensor-based graph cuts, and Voronoi diagrams estimates renal vascular dominant regions with 80% Dice on 8 cases.","lead":"Researchers built an automatic pipeline that segments kidneys and renal arteries in CT scans, then uses Voronoi territory maps to show which artery branch feeds which part of the kidney. The goal is pre-surgical planning for partial nephrectomy, where surgeons clamp only the blood vessels that supply the tumor.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 80% Dice for dominant-region estimation is computed against a simulated ground truth built with the same Voronoi algorithm, so it measures segmentation consistency rather than anatomical precision; the central claim of precise estimation remains unvalidated.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue: the dominant-region Dice is computed against a reference produced by the same Voronoi algorithm applied to manual segmentations, so it cannot validate the biological fidelity of the Voronoi model. My review confirms this and adds that the renal artery segmentation step, whose branching structure directly determines the Voronoi partition, is not re-evaluated on the 8 test cases, compounding the uncertainty. The paper is honest about not having anatomical ground truth and presents a coherent pipeline with careful 8-fold cross-validation for kidney segmentation; these strengths support a conditional acceptance with the requirement that the central estimation claim be tested against an independent measure of perfusion. The reader's CONDITIONAL verdict is therefore appropriate, and my concern does not change it.","tokens_in":14699,"tokens_out":5636,"duration_ms":56016,"concrete_test":"Perform a controlled intraoperative validation in 5-10 renal units: clamp a single segmental renal artery, inject dye or use indocyanine green fluorescence to delineate the ischemic/stained perfusion territory, and compare this dye-defined territory with the Voronoi-predicted dominant region from the automatic pipeline using per-zone Dice. If the median Dice against dye-defined territories is substantially below the 80% reported against the simulated reference, the reported accuracy is a self-consistency artifact rather than evidence of anatomical precision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 states that anatomical ground truth for dominant regions is unavailable, so the reference is generated by applying the same Voronoi algorithm (Eq. 5) to manual kidney and artery segmentations. The Dice score in Table 2 therefore quantifies how closely the automatic pipeline reproduces the manual inputs under that algorithm, not whether Euclidean-distance Voronoi zones correspond to real blood-perfusion territories. The only external evidence is a single surgical case in Section 5.3: four regions were predicted adjacent to the tumor, surgeons clamped two, and slight bleeding remained in the two unclamped predicted regions. This is anecdotal and does not establish quantitative accuracy. Moreover, renal artery segmentation, the main geometric driver of the Voronoi partition, is not re-measured on the evaluation cases (Section 3), and the paper shows under-segmented arteries in Cases 2 and 6 (Fig. 16) that change the dominant-region estimates. Thus the central claim of precise estimation rests on an unvalidated equivalence between Voronoi proximity and true arterial supply.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully automatic pipeline for estimating renal vascular dominant regions from contrast-enhanced CT: a spatially aware 3D U-Net-like FCN for kidney segmentation, a tensor-based graph-cut method (tensor-cut) for renal artery segmentation, and a Voronoi diagram (Eq. 5) to partition the kidney into regions attributed to each arterial branch. Kidney segmentation is evaluated on 27 cases with 8-fold cross-validation against U-Net and V-Net baselines. Dominant-region estimation is evaluated in 8 cases by comparing automatic Voronoi partitions against a simulated ground truth built by applying the same Voronoi algorithm to manual kidney and artery segmentations. A single clinical selective-clamping case is presented as an illustration. The authors position the work as a preliminary study and explicitly note that anatomical ground truth for dominant regions is unavailable.","tokens_in":14961,"tokens_out":2613,"duration_ms":27086,"significance":"If the claims were fully supported, the pipeline would be a useful engineering contribution to pre-surgical planning for partial nephrectomy: it is fully automatic, combines well-established components in a sensible way, and the kidney segmentation evaluation is solid for a small in-house dataset, including 8-fold CV and comparisons to two standard baselines. The paper is honest about several limitations, including the lack of anatomical ground truth for dominant regions and the absence of a new quantitative evaluation of renal artery segmentation. However, the central claim of 'precise estimation' of dominant regions is not established by the reported experiments. The 80% Dice figure measures agreement with a simulated reference produced by the same algorithm applied to manual inputs, not accuracy against any independent or anatomical standard. The paper's value currently lies in demonstrating pipeline feasibility, not in validating the Voronoi model as a representation of true perfusion territories.","major_comments":[{"comment":"The abstract claims 'The accuracy of kidney segmentation in 27 cases with 8-fold cross validation reached a Dice score of 95%.' Table 1 reports for the proposed method a left-kidney median DSC of 87.3% and a right-kidney median DSC of 94.7%, with no reported overall 95% figure. The abstract's number appears to be the right-kidney median or a misreading; please state a single overall metric or report all values consistently, and adjust the abstract to match the results table.","section":"Abstract & §4.1, Table 1"},{"comment":"Section 3 states that 'we did not perform a quantitative evaluation for blood vessel segmentation in this work,' and §4.2 only cites the previous tensor-cut paper (Ref. [44]) for a centerline overlap exceeding 80%. Because the Voronoi partition in Eq. (5) depends directly on the arterial branches B, the accuracy of the dominant-region estimation is load-bearing on a component that is not re-measured on the evaluation cases. The manuscript should either re-measure renal artery segmentation on the 8 cases used for dominant-region evaluation or explicitly quantify how segmentation errors propagate to the Voronoi cells.","section":"§3 and §4.2"},{"comment":"The quantitative evaluation of dominant-region estimation is circular for validating the Voronoi model. As stated in §4.3, the 'simulated ground truth' is generated by applying the same Voronoi algorithm (Eq. 5) to manual kidney and artery segmentations. Consequently, the Dice coefficients in Table 2 measure how closely the automatic pipeline reproduces the manual inputs under the Euclidean-distance Voronoi rule, not whether that rule corresponds to true renal perfusion territories. The text should be rewritten so that the 80% Dice is presented as a reproducibility/consistency measure, not as 'estimation accuracy.' The single surgical case in §5.3 is anecdotal and does not provide quantitative validation.","section":"§4.3, Eq. (5), and Table 2"},{"comment":"The discussion acknowledges under-segmented renal arteries in Cases 2 and 6 and states that for Case 6 'the blue region is over-estimated.' Since these under-segmented arteries directly alter the Voronoi partition, the claim in the same section that 'the renal artery segmentation performance for dominant-region estimation is acceptable in this work' is not supported by quantitative evidence. Please report the effect of segmentation errors on the dominant-region Dice for these and other cases, or temper the claim to reflect that the impact was assessed only qualitatively.","section":"§5.2 and Fig. 16"}],"minor_comments":[{"comment":"The sentence 'As shown in Table 2, compared with other U-Net-like architectures...' appears to refer to Table 1, which contains the method comparison; Table 2 reports dominant-region Dice. Please correct the reference.","section":"§5.1"},{"comment":"The column header 'Area(mm3)' should be 'Area(mm²)' since the contact area is a two-dimensional quantity.","section":"Table 3"},{"comment":"The Voronoi partition depends on a branch-clustering level, but the paper does not specify how this level is selected for the reported experiments. Since different clustering levels produce different partitions (Fig. 6), please state the criterion used in the evaluation and whether any parameters were tuned.","section":"§2.3, Eq. (5)"},{"comment":"The captions state 'Validation DSC' for both pre-training and fine-tuning curves and for the comparison plot. In the comparison plot, it would help to state explicitly that the curves show median validation DSC across the 8 folds, with quartile bands, in the caption rather than only in the main text.","section":"Fig. 8 and Fig. 9"},{"comment":"The sentence 'The vesselness enhancement filter transfers the higher dimensional tensor to 1-dimensional Euclidean measurement V' is grammatically unclear and should be rephrased.","section":"§2.2"},{"comment":"The layout of Table 2 is hard to read: the 'Kidney #' and 'Regions #' rows are typeset as a single line in the supplied text, and the parentheses for Dice values are unbalanced. Please reformat the table so each row is clearly separated.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the fully automatic pipeline: a spatially aware 3D U-Net for the kidney, their prior tensor-cut for arteries, and a Voronoi step to get vascular territories. The spatial-coordinate trick at the bottom feature map is a small, sensible engineering change, and the 8-fold cross-validation on 27 cases is done properly, with U-Net and V-Net baselines under the same protocol. On those terms the kidney segmentation result is credible: the spatial unit reduces false positives and improves Dice, and the authors are honest about the two failed cases with 2.0 mm slice thickness.\n\nThe soft spots are in the dominant-region evaluation, and they are real. The abstract says the final Dice is 80%, but that number is computed against a simulated ground truth: they run the same Voronoi algorithm on manual kidney and artery segmentations and treat that as the reference. So the Dice measures how closely the automatic pipeline reproduces the manual inputs under that algorithm, not whether Euclidean Voronoi zones correspond to real blood perfusion. The paper acknowledges this in Section 4.3, but the abstract does not, and the word 'precise' in the title overstates what is validated. The renal artery segmentation itself is not re-measured on the evaluation cases; they rely on the prior paper's 80% centerline overlap. Given that under-segmented arteries directly change the Voronoi boundaries (they show this in Cases 2 and 6), that omission matters. The single surgical case is anecdotal: four predicted adjacent regions, two clamped, slight bleeding in the others. That is promising but not quantitative validation.\n\nAlso minor: the abstract's '95% Dice' for kidney segmentation is inconsistent with Table 1, which reports 87.3% for the left kidney and 94.7% for the right. That should be fixed before publication.\n\nI do think the central argument holds up if you read it as a feasibility study: the components are known, the integration is real, and the limitations are stated in the discussion. The paper is worth a serious referee, but the referee should push for re-framing the dominant-region results as segmentation-consistency rather than anatomical accuracy, and for at least a sensitivity analysis of the Voronoi assumption on a few cases with actual perfusion data.\n\nMy recommendation: engage with it. It is a useful engineering contribution for partial nephrectomy planning, and the evaluation can be made honest with relatively modest revisions.","headline":"A coherent fully automatic pipeline for renal vascular dominant-region estimation, with solid kidney-segmentation experiments, but the headline 80% accuracy for dominant regions is consistency with a Voronoi simulation, not anatomical truth.","tokens_in":15473,"tokens_out":1043,"would_cite":true,"duration_ms":13251,"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":"Fully automatic kidney-territory mapping reaches 80% Dice.","keywords":["kidney segmentation","renal artery segmentation","Voronoi diagram","vascular dominant region","partial nephrectomy","fully convolutional networks","tensor-cut","graph cut"],"falsifier":"Compare the predicted dominant regions against observed perfusion in the same patients, for example using CT perfusion maps, selective arterial dye injection, or indocyanine-green fluorescence during surgery. If the boundaries between territories shift by more than a clinically meaningful margin, such as more than 10 mm, in a substantial fraction of cases, the Euclidean-distance Voronoi assumption, not the segmentations, would be the thing that fails.","tokens_in":14520,"feed_emoji":"🗺️","tokens_out":4845,"duration_ms":45242,"temperature":0.7,"pith_summary":"This paper reports a fully automatic pipeline that turns a contrast-enhanced CT scan into a map of which renal artery supplies which part of the kidney. The goal is to help surgeons plan partial nephrectomy by showing which blood vessels feed a tumor before any clamping decision is made. The pipeline combines a spatially aware 3D fully convolutional network for kidney segmentation, a tensor-based graph-cut method called tensor-cut for renal artery extraction, and a Voronoi diagram that partitions the kidney into vascular dominant regions. On 27 cases the kidney segmentation reached a Dice score of 95%; on 8 cases the artery segmentation reached an 80% centerline overlap and the final dominant-region estimation reached a Dice score of 80%. A sympathetic reader would take the paper's claim as: the components work together end-to-end, and the main remaining validation is clinical, on larger data.","feed_headline":"Fully automatic kidney-territory mapping reaches 80% Dice.","feed_subtitle":"From a single CT scan, the pipeline tells surgeons which kidney arteries to clamp around a tumor.","key_machinery":"The load-bearing mechanism is the Voronoi diagram of Eq. 5: each branch $B_i$ of the renal artery tree acts as a set of seed points, and every voxel $x$ inside the segmented kidney $X_v$ is assigned to the branch whose minimal Euclidean distance $d(x,B_i)$ is smallest, so each Voronoi cell is the simulated perfusion territory of one artery branch. Around that core, the pipeline uses two supporting mechanisms: a spatially aware 3D fully convolutional network that concatenates normalized $x,y,z$ coordinate channels to the bottom feature map so cropped sub-volumes retain their position in the whole CT, and tensor-cut, which models the Hessian matrix as a second-order tensor, measures tensor distances with an affine-invariant Riemannian metric, and solves a first-order Markov random field by graph cut. The Voronoi construction is also applied at different branch-clustering levels to produce coarse-to-fine partitions.","core_discovery":"The paper claims that renal vascular dominant regions can be estimated precisely and automatically by treating each renal artery branch as a seed set in a Voronoi diagram computed inside the segmented kidney. For the inputs to that diagram, the paper argues that a 3D U-Net with normalized x, y, z coordinate maps concatenated at the bottom feature map segments kidneys more accurately than plain U-Net or V-Net baselines, and that the tensor-cut method segments tiny, low-contrast renal arteries where Hessian vesselness filters fail. The reported end-to-end accuracy is a Dice coefficient of 80% for dominant regions, measured against simulated ground truth built by applying the same Voronoi rule to manual kidney and artery segmentations. The paper also reports that in one clinical case, surgeons used the predicted dominant regions to clamp two of four tumor-adjacent arteries, with slight residual bleeding in two regions that were not clamped.","pith_inferences":["A testable prediction is that most of the remaining 20% Dice gap is attributable to under-segmentation of small artery branches near the tumor, not to the Voronoi rule itself; the paper's Cases 2 and 6 point in that direction.","The method could be extended to other organs with segmental blood supply, such as liver segments or lung lobes, by swapping the FCN and vessel segmentation but keeping the Voronoi partition.","Because the reference is algorithmically generated, the clinical value of the 80% number depends on a future perfusion-based validation; without that, the metric should be read as an error-propagation study, not an anatomical accuracy claim.","An immediate extension would be to use the contact-area ratio, such as region 5 with 78% of tumor-adjacent area in the example, as the clamping priority score rather than region volume."],"forward_implications":["Selective clamping in partial nephrectomy could be planned from a CT alone, without manual kidney or artery tracing.","Adding the coordinate map to a 3D U-Net appears to suppress false positives in kidney segmentation compared with the same network without it.","Tiny, low-contrast artery branches that Hessian vesselness filters miss can be recovered by tensor-cut, which is required for fine Voronoi partitions.","Multi-level Voronoi partitions give surgeons both coarse and fine views of which artery branches dominate the region around a tumor.","The reported Dice of 80% sets a quantitative baseline that future end-to-end pipelines for vascular-territory estimation can be compared against."],"supporting_citations":[{"why":"Supplies the tensor-cut renal artery segmentation method used directly in this pipeline.","marker":"[44]"},{"why":"Provides the U-Net-like architecture and pre-trained baseline that the spatially aware FCN builds on and compares against.","marker":"[31]"},{"why":"V-Net is implemented as a comparison baseline on the same kidney dataset.","marker":"[26]"},{"why":"Earlier clinical work showing Voronoi diagrams applied to renal dominant regions for partial nephrectomy planning.","marker":"[42]"},{"why":"Prior virtual partial nephrectomy analysis that uses Voronoi diagrams to estimate dominant regions.","marker":"[21]"},{"why":"Prior computational robot-assisted partial nephrectomy planning that also relies on Voronoi-based dominant regions.","marker":"[19]"},{"why":"Frangi vesselness enhancement filter, the conventional approach that tensor-cut is designed to improve upon.","marker":"[11]"},{"why":"Supplies the affine-invariant Riemannian metric used to measure tensor distances in tensor-cut.","marker":"[29]"},{"why":"Inspiration for incorporating spatial coordinate information into the encoder of the network.","marker":"[6]"},{"why":"Multi-organ segmentation model used for pre-training before fine-tuning on the kidney dataset.","marker":"[39]"}],"fun_headline_variants":["Automatic kidney territory mapping hits 80% Dice","Voronoi maps kidney blood supply for surgical planning","AI-driven renal territory estimation reaches 80% accuracy","Kidney artery territories auto-mapped for clamping decisions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scoring of the 80% accuracy uses simulated ground truth: the same Voronoi algorithm is applied to manual kidney and artery segmentations, so the Dice number measures the cost of automatic segmentation errors, and the whole approach presupposes that Euclidean-distance Voronoi zones are an acceptable model of real blood perfusion territories.","fun_headline_variants_meta":{"raw":{"variants":["Automatic kidney territory mapping hits 80% Dice","Voronoi maps kidney blood supply for surgical planning","AI-driven renal territory estimation reaches 80% accuracy","Kidney artery territories auto-mapped for clamping decisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1183,"prompt_tokens":943,"completion_tokens":240,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":177}},"tokens_in":559,"tokens_out":240,"duration_ms":3138,"temperature":1.0,"reasoning_tokens":177,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:09:23.458524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the predicted dominant regions against observed perfusion in the same patients, for example using CT perfusion maps, selective arterial dye injection, or indocyanine-green fluorescence during surgery. If the boundaries between territories shift by more than a clinically meaningful margin, such as more than 10 mm, in a substantial fraction of cases, the Euclidean-distance Voronoi assumption, not the segmentations, would be the thing that fails.","supporting_citations":[{"cited_title":"F., Mori, K., 2016","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor-cut renal artery segmentation method used directly in this pipeline."},{"cited_title":"Towards dense volumetric pancreas segmentation in CT using 3d fully convolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides the U-Net-like architecture and pre-trained baseline that the spatially aware FCN builds on and compares against."},{"cited_title":"V-net: Fully convolu- tional neural networks for volumetric medical image segmentation","cited_arxiv_id":null,"evidence_quote":"V-Net is implemented as a comparison baseline on the same kidney dataset."},{"cited_title":"S., 2012","cited_arxiv_id":null,"evidence_quote":"Earlier clinical work showing Voronoi diagrams applied to renal dominant regions for partial nephrectomy planning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior virtual partial nephrectomy analysis that uses Voronoi diagrams to estimate dominant regions."},{"cited_title":"Feasibility and accuracy of computational robot-assisted partial nephrectomy planning by virtual partial nephrectomy analysis","cited_arxiv_id":null,"evidence_quote":"Prior computational robot-assisted partial nephrectomy planning that also relies on Voronoi-based dominant regions."},{"cited_title":"F., Niessen, W","cited_arxiv_id":null,"evidence_quote":"Frangi vesselness enhancement filter, the conventional approach that tensor-cut is designed to improve upon."},{"cited_title":"A riemannian framework for tensor computing","cited_arxiv_id":null,"evidence_quote":"Supplies the affine-invariant Riemannian metric used to measure tensor distances in tensor-cut."},{"cited_title":"S., Hwang, J.-N., Yuan, C., 2017","cited_arxiv_id":null,"evidence_quote":"Inspiration for incorporating spatial coordinate information into the encoder of the network."},{"cited_title":"On the influence of Dice loss function in multi-class organ segmentation of abdominal CT using 3D fully convolutional networks","cited_arxiv_id":"1801.05912","evidence_quote":"Multi-organ segmentation model used for pre-training before fine-tuning on the kidney dataset."}],"review_version":1}