{"id":"edaf1ed2-b874-4a34-87b2-d0186c190e43","arxiv_id":"2505.21644","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Ridge-guided point prompts let SAM segment more plant root pixels with fewer prompt points than uniform grid prompts on minirhizotron images.","lead":"This paper pairs the Segment Anything model with a geometric ridge detector that places prompt points on long, bright, root-like structures in minirhizotron images. It reports that focused prompting captures more root pixels per prompt than uniform grid points, which could reduce manual labeling effort in plant root studies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported TPR/FPR comparison rests on benchmark masks that the authors themselves describe as incomplete; the geomprompt vs grid advantage at low prompt counts may be an artifact of reference-annotation error.","rationale":"Good-faith reading: the paper is an engineering proposal, not a claimed theoretical advance, and the code is released. The most defensible version of the central claim is comparative: geometric prompts make SAM more sample-efficient than grid prompts on root images. That comparison is exactly what Table II quantifies, so the validity of the reference masks is the load-bearing condition. The reader flagged this too, and the paper's own text makes it inescapable: Section III-B documents incomplete annotations; Section III-D2 admits the objective mismatch. The other weaknesses (unreported ridge tuning parameters, lack of random baseline, no confidence intervals) are real but would be of secondary importance; a clean reference is required even to interpret them. Machine-checked proofs and reproducible code are present, but they cannot compensate for an unreliable ground-truth benchmark. Thus I do not propose rejecting the paper; I propose that the conditional verdict stands pending the clean-reference check. I agree with the reader's weakest-assumption identification.","tokens_in":10527,"tokens_out":3055,"duration_ms":32932,"concrete_test":"Select a stratified random subset of roughly 100 images from the 2,419 switchgrass images. Have two independent annotators trace all visible roots, including faint ones, using a written protocol, then adjudicate disagreements to produce a 'clean' reference. Recompute Table II TPR/FPR/IoU for geomprompt and grid at 16/64/256/1024 prompts against this clean reference, reporting per-image differences and paired confidence intervals. If the 64-prompt TPR margin (geomprompt minus grid) and FPR margin remain materially positive and statistically distinguishable from zero, the claim is supported; if the margins collapse or reverse, the reported advantages were an artifact of annotation incompleteness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B states that 'pixel level annotations of root features are frequently inaccurate' and that 'features that appear to be roots are but are not annotated'; Section III-D2 concedes 'an objective mismatch' between segmenting ridge-like features and segmenting roots of interest. Despite this, Table II's headline numbers (77.6% TPR / 16.7% FPR for 64 geomprompt vs 51.9% / 22.1% for 64 grid) are computed pixel-wise against these same benchmark masks. If the reference masks omit true roots, then geomprompt's tendency to segment ridge-like features that are in fact roots will be scored as false positives, inflating its FPR, while a method that ignores those features could appear spuriously specific. Conversely, TPR is bounded by the reference's completeness rather than by true sensitivity. Because the authors themselves supply evidence that the reference is incomplete, the central empirical claim—that geomprompt is more sensitive and specific at low prompt counts—is not yet established. The absence of per-image or variance statistics on the aggregated 2,419-image comparison makes it impossible to see how much of the 25.7-point TPR gap at 64 prompts depends on which images dominate the aggregate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeomPrompt, a non-learned prompt generator that places SAM point prompts on multiscale ridge features, targeting root segmentation in minirhizotron images. It compares geomprompt against uniform grid prompting at 16, 64, 256, and 1024 points on 2,419 switchgrass images from the PRMI dataset, reporting higher TPR and lower FPR at low prompt densities (e.g., 77.6% vs 51.9% TPR at 64 points). The authors also release the geomprompt Python package and emphasize that SAM is used out of the box with no fine-tuning.","tokens_in":10682,"tokens_out":4878,"duration_ms":52373,"significance":"If validated, the result is practically significant: it would enable zero-shot SAM-based root segmentation with only a few dozen prompts, substantially reducing the annotation burden in rhizotron studies. The release of open-source software and the use of a stock SAM checkpoint are notable strengths, as is the general idea of using non-learned geometric prompts to focus a foundation model's attention. However, the empirical evidence as presented is not yet sufficient to support the central claim of superiority at low prompt counts, because the evaluation relies on imperfect reference masks and lacks uncertainty quantification.","major_comments":[{"comment":"The reference masks are acknowledged in Section III-B as 'frequently inaccurate' with unannotated root-like features, and Section III-D2 concedes an 'objective mismatch' between segmenting ridge-like features and segmenting roots of interest. The headline TPR/FPR numbers in Table II are computed pixel-wise against these same masks. Since geomprompt is designed to detect ridge-like features that may include unannotated roots, its FPR could be inflated relative to a method that avoids those regions, and its TPR ceiling is determined by annotation completeness. The authors should either evaluate on a subset of images with carefully corrected reference masks, or conduct a sensitivity analysis that treats obvious unannotated roots as true positives and shows how the reported advantage changes. As written, the central empirical claim rests on an unvalidated ground truth.","section":"Section III-B and III-D2, Table II"},{"comment":"The aggregated TPR/FPR values across 2,419 images are reported without confidence intervals, standard errors, or significance tests. The 25.7-point TPR gap at 64 prompts could be driven by a small number of influential images. The authors should report per-image distributions or bootstrap confidence intervals and a paired statistical test (e.g., Wilcoxon signed-rank) for each prompt density to demonstrate that the observed differences are stable and not due to outliers.","section":"Table II"},{"comment":"The ridge detector's scale range and step (the 'series of K scales' in Section II-C), the thresholds used to extract connected ridge curves, and the random sampling of points within each curve are not reported, and no random seed is stated. The authors say in Section III-C that they 'tuned the ridge detection algorithm to obtain an approximately correct number of geomprompt points at each scale,' but the tuning parameters are absent. Without these details, the comparison is not reproducible, and it is unclear whether the prompt counts are truly comparable (e.g., 64 geomprompt vs 64 grid points) in terms of algorithmic effort.","section":"Section III-A and III-C"},{"comment":"The abstract and introduction promise a comparison with 'a custom algorithmic segmentation routine,' but Section III and Tables I-II only report geomprompt versus grid prompting. Additionally, Section II-B asserts that geomprompt 'outputs segmentations that tend to be more accurate than benchmark hand annotations,' a claim not supported by the reported metrics, which show low IoU and relatively high FPR. Please either remove these statements or supply the corresponding experiments and supporting evidence.","section":"Section I and II-B"}],"minor_comments":[{"comment":"There is a typo: 'withinin situimages' should be 'within in situ images'; also 'segmention' in the introduction should be 'segmentation.'","section":"Abstract"},{"comment":"The caption reads 'an the accompanying annotation' and should be 'and the accompanying annotation.'","section":"Figure 2 caption"},{"comment":"The caption contains a typo: 'GEOMPPROMPT' should be 'GEOMPROMPT.'","section":"Table II caption"},{"comment":"The text says '74% of geomprompt points' and '47% of grid prompt points' at 16 prompts, while Table I lists 73.9% and 46.8%; please reconcile the rounding.","section":"Section III-D1"},{"comment":"The claim that focused prompting will yield 'much more important efficiency gains in the application of human and expert attention' is plausible but presented without a user study; consider softening or citing evidence from active learning literature.","section":"Section IV"},{"comment":"The sentence 'we posit that random sampling of prompt points from a 2-D uniform distribution over the image space would result in comparable segmentation performance' is an unsupported conjecture; it can be removed or tested by including a random-prompt baseline.","section":"Section III-D1"}],"recommendation":"major_revision","confidential_remarks":"The paper describes a practically motivated method and ships an open-source package, but the evaluation as it stands is not yet at the standard expected for a journal submission. The authors' own admission of imperfect ground truth, combined with the absence of uncertainty quantification and unreported tuning parameters, makes the central claim difficult to assess. Major revision is appropriate: the required additions are feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth sending to a serious referee. What's new is a simple, training-free recipe: run Lindeberg multiscale ridge detection, rank connected ridge curves by salience, allocate SAM point prompts proportionally, and segment. On 2,419 switchgrass minirhizotron images they report that this beats uniform grid prompting at low prompt counts (77.6% TPR / 16.7% FPR at 64 prompts vs 51.9% / 22.1% for grid). Code is released, the writing is clear, and the authors are upfront that the PRMI benchmark masks are imperfect.\n\nThe soft spots are real but not fatal. The headline comparison is aggregated without confidence intervals or significance tests, so we can't tell whether the gap is driven by a few images. Ridge-detection and SAM threshold parameters are tuned but the tuning values aren't fully reported. There's no random-prompt baseline, though the authors correctly note that random uniform prompts should behave like the grid.\n\nOn the bigger concern about ground truth: the stress-test worry that the geomprompt advantage might be an artifact of annotation error doesn't actually land. The authors' own admission cuts both ways. If the reference masks omit true roots, a method that finds those roots gets penalized with false positives and misses out on true positives, while a method that ignores them looks spuriously specific. That would make geomprompt's measured advantage conservative, not inflated. The deeper mismatch is the paper's explicit 'objective mismatch' between segmenting ridge-like features and segmenting roots of interest, which limits the practical claim and should be addressed head-on.\n\nWho is this for? People in root phenotyping, and more broadly anyone thinking about prompt engineering for foundation models in scientific imaging. It's a niche but useful engineering result, not a conceptual breakthrough. The evaluation is tractable and reproducible thanks to the code release. I'd send it to review with a request for variance statistics, a random-prompt baseline, the exact ridge parameters, and ideally one additional species or soil type. The central idea is plausible and the paper is honest about its limitations.","headline":"A solid, niche engineering paper: ridge-salience point prompting of SAM beats grid prompting for root segmentation, with honest limitations and released code; the missing statistics and imperfect ground truth warrant a careful revision before acceptance.","tokens_in":11321,"tokens_out":3300,"would_cite":false,"duration_ms":36441,"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":"Prompting the Segment Anything Model with points placed on multiscale ridge features produces more sensitive and more specific root segmentations than uniform grid prompting at low prompt counts, without fine-tuning the model.","keywords":["segment anything model","point prompting","multiscale ridge detection","root segmentation","minirhizotron imagery","foundation models","geometric features","scale-space analysis"],"falsifier":"Compute geomprompt and grid-prompted SAM TPR/FPR on a rhizotron test set whose root masks are verified by multiple independent annotators (or generated synthetically from known root geometry). If the true-positive gap at 64 prompts shrinks to near zero or the false-positive rates invert once unmasked roots are counted, then the reported advantage is an artifact of incomplete ground truth.","tokens_in":10246,"feed_emoji":"🌱","tokens_out":5914,"duration_ms":53894,"temperature":0.7,"pith_summary":"The paper claims that the Segment Anything Model (SAM), prompted by points automatically placed on multiscale ridge features in an image, can segment fine plant roots from soil backgrounds with fewer prompts and better accuracy than SAM prompted with an equally dense uniform grid of points. The method, called geomprompt, uses classical ridge detection from scale-space theory to locate elongated bright structures, then feeds those locations as point prompts to SAM out of the box, without any fine-tuning. On a benchmark of 2,419 switchgrass minirhizotron images, 64 geomprompt points capture 77.6% of annotated root pixels at a 16.7% false positive rate, versus 51.9% and 22.1% for 64 grid points. A sympathetic reader would care because rhizotron root segmentation has historically been a manual, subjective, and labor-intensive bottleneck in plant science, and this result points to a cheap automatic alternative using an existing foundation model.","feed_headline":"Ridge-point prompts lift SAM's root recall from 52% to 78%","feed_subtitle":"Prompting Segment Anything with multiscale ridge features finds fine roots without fine-tuning.","key_machinery":"The central object is the multiscale ridge detector with gamma-normalized principal curvature difference as the ridge strength measure, combined with a ridge-salience-weighted prompt allocation scheme. Ridge curves are detected in scale-space as connected components whose ridge test value is integrated along the curve into a salience score; a budget of K prompts is then distributed across curves in proportion to their salience. This object carries the argument because it converts the geometric prior 'roots are bright and elongated' into a small set of point locations that SAM can use directly as prompts, and the proportional allocation avoids both the clustering of prompts at high-intensity blobs and their wasteful dispersal across random soil features.","core_discovery":"The central discovery is that the relationship between prompts and segment quality in SAM is not fixed: prompting the model at locations that carry a geometric prior about the target feature concentrates the model's attention on that feature class. Using Lindeberg's multiscale ridge detector, the authors generate prompt points on ridge-like image structures, which in rhizotron images correspond to roots. At low prompt densities this focused prompting outperforms grid prompting substantially on pixel-level recall and false positives, and the advantage only disappears at roughly 1024 prompts per image, where grid prompting matches geomprompt's true positive rate. The paper further reports that segmented root structures are instance-level, which matters for non-destructive studies of root growth and turnover, and that the approach generalizes in principle to SAM 2 and to other ridge-like or valley-like structures.","pith_inferences":["Because the paper itself notes the benchmark masks miss many root-like features, geomprompt's reported false positive rate is likely inflated relative to a more complete ground truth; a fairer benchmark could widen rather than shrink the demonstrated advantage.","The recipe generalizes beyond roots: any point-prompted foundation model can be steered by classical geometric priors (edges, blobs, ridges, valleys) at negligible cost, offering an alternative to learned prompters trained for a single domain.","A direct testable extension would run the same comparison on SAM 2 over minirhizotron video frames: if ridge prompts add temporal consistency, the approach could support automated root dynamics tracking without further training.","The method's sensitivity to ridge-detection parameters (scale range, gamma, and salience thresholds) across heterogeneous soil types and root species is untested; that sensitivity is the likeliest source of failure in new experiments."],"forward_implications":["SAM, prompted by geomprompt, can produce usable root segmentations without fine-tuning, potentially replacing hand tracing as the first pass in rhizotron experiments.","Instance-level root segments become available automatically, enabling non-destructive studies of individual root growth and turnover.","Fewer prompts mean less post-processing for overlapping-mask disambiguation in SAM's segment-everything mode, offsetting some of the ridge-detection overhead.","The method transfers to other segmentation targets with ridged or tubular morphology, such as blood vessels, and to valley-like dark structures by inverting the ridge test.","Fine-tuning a check-pointed SAM on geomprompt outputs, with light filtering, is a plausible route to further accuracy gains beyond the out-of-the-box results reported here."],"supporting_citations":[{"why":"The Segment Anything Model that is prompted; this is the segmentation engine whose prompt sensitivity the paper exploits.","marker":"[9]"},{"why":"The multiscale ridge detection theory that supplies the geometric features from which prompt points are derived.","marker":"[10]"},{"why":"The PRMI minirhizotron dataset that provides the switchgrass images and benchmark masks used in the TPR/FPR evaluation.","marker":"[35]"}],"fun_headline_variants":["Geometric prompts boost SAM root segmentation accuracy","Ridge-based prompts improve SAM's root detection in rhizotron images","Prompting SAM with ridge features finds fine roots without fine-tuning","New prompt strategy lifts SAM root recall from 52% to 78%","Multiscale ridge prompts sharpen SAM's root segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation relies on the PRMI switchgrass pixel masks as ground truth even though the paper observes that many root-like features are not annotated and some annotations are inaccurate, so the reported true-positive and false-positive rates could be biased if those masks are systematically incomplete.","fun_headline_variants_meta":{"raw":{"variants":["Geometric prompts boost SAM root segmentation accuracy","Ridge-based prompts improve SAM's root detection in rhizotron images","Prompting SAM with ridge features finds fine roots without fine-tuning","New prompt strategy lifts SAM root recall from 52% to 78%","Multiscale ridge prompts sharpen SAM's root segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1365,"prompt_tokens":930,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":351}},"tokens_in":546,"tokens_out":435,"duration_ms":4594,"temperature":1.0,"reasoning_tokens":351,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:25:53.259593+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute geomprompt and grid-prompted SAM TPR/FPR on a rhizotron test set whose root masks are verified by multiple independent annotators (or generated synthetically from known root geometry). If the true-positive gap at 64 prompts shrinks to near zero or the false-positive rates invert once unmasked roots are counted, then the reported advantage is an artifact of incomplete ground truth.","supporting_citations":[{"cited_title":"Semantic segmentation of plant roots from rgb (mini-) rhizotron images—generalisation potential and false positives of established methods and advanced deep-learning models,","cited_arxiv_id":null,"evidence_quote":"The PRMI minirhizotron dataset that provides the switchgrass images and benchmark masks used in the TPR/FPR evaluation."}],"review_version":1}