{"id":"98ff599a-628b-4b57-9fb1-3836927d8beb","arxiv_id":"2501.05236","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Texture features plus a support vector machine can outline external cervical resorption in CBCT scans, but the three-patient evaluation with test-set tuning leaves the reported accuracy unproven.","lead":"Dentists need to track a tooth condition where cells eat away the root, and the paper tests whether texture patterns in 3D CT scans can automatically outline these lesions. If larger studies confirm this, dentists could monitor healing or progression without slow, subjective manual tracing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Dice is selected from the same leave-one-out results used to report it, so the central accuracy claim rests on test-set tuning; an independent-patient selection protocol is needed.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: feature and hyperparameter selection are performed on the same data used to report performance, with a three-patient cohort. This is the most serious threat to the central claim because the reported Dice is the only quantitative evidence for accurate detection. The paper itself acknowledges the small sample size but does not disclose or mitigate the selection-on-test-data issue. My proposed concrete test—nested leave-one-out with selection restricted to training patients—would settle whether the reported accuracy survives an honest evaluation. Since the reader already recommended CONDITIONAL and my concern does not push the verdict to REJECT (the method may still work after proper validation), the verdict remains CONDITIONAL; no change is needed.","tokens_in":136,"tokens_out":1460,"duration_ms":22875,"concrete_test":"Re-run the evaluation with a nested leave-one-out protocol: for each held-out patient, perform feature selection (LGRE/HGRE) and radius selection (2/5/7/9) using only the two training patients' data—e.g., by internal leave-one-out within those two, or a simple validation split—then train the SVM with those choices and evaluate on the held-out patient. Report the resulting Dice, precision, and recall per patient and the mean. If the mean Dice falls below ~0.5 or the per-patient spread widens substantially, the original Table 1 results reflect test-set tuning rather than generalizable accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that LGRE and HGRE texture features accurately detect ECR in CBCT—is supported only by a leave-one-out cross-validation in which the test patients are used to make the very choices that determine the reported score. In Section 3.2, LGRE and HGRE were selected by visually comparing feature histograms on the entire dataset, including the test patient's voxels. In Section 4, the neighborhood radius (2, 5, 7, 9) was chosen as 5 because it gave the best Dice in Table 1 on the same leave-one-out folds. This is test-set tuning: every fold's evaluation is influenced by global feature selection and the final radius selection. With only three patients, the risk of optimistic bias is substantial; the average Dice of 0.59, with one case performing poorly and another degrading sharply at radius 9, shows instability. Without a clean separation between model selection and evaluation, the reported accuracy cannot be taken as an unbiased estimate of performance on new patients, and the abstract's claim of accurate detection is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an automated method for segmenting external cervical resorption (ECR) lesions in cone-beam CT (CBCT) scans. The method extracts voxel-wise texture features (GLCM and GLRLM based) from cropped tooth regions, selects two features (LGRE and HGRE) by visual histogram comparison, and trains a linear SVM classifier for binary lesion/healthy classification. Evaluation is performed with leave-one-out cross-validation at the patient level on six scans from three patients, reporting an average Dice score of 0.59. The authors also present an unsupervised k-means clustering analysis of within-lesion texture features for one case to identify possible calcification patterns. The central claim is that certain texture features can accurately detect subtle CBCT signal changes due to ECR.","tokens_in":4380,"tokens_out":4198,"duration_ms":39252,"significance":"If the accuracy claim were properly supported, this work would be a valuable step toward automated ECR assessment, addressing a real clinical need for objective, reproducible lesion quantification in CBCT. The use of interpretable texture features, reliance on an open-source ITK module, and clear description of the processing pipeline are strengths. However, the current evaluation is compromised by model selection on the test data and the very small sample size, so the reported performance cannot be taken as evidence for generalization to new patients. The clinical problem is important and the method is clearly described, but the central claim goes beyond what the evidence supports.","major_comments":[{"comment":"The model selection procedure uses the test data in two ways. First, Section 3.2 states that LGRE and HGRE were selected by visually examining histograms of feature distributions computed over the entire dataset; in every leave-one-out fold, the 'held-out' patient's voxels were thus used to choose the features. Second, Section 4 states that the neighborhood radius of 5 voxels was selected based on the leave-one-out Dice results in Table 1. Consequently, the reported DSC of 0.59 is not an unbiased estimate of performance on unseen patients; it is the outcome of fitting both the feature set and a free parameter to the same cases on which accuracy is reported. With only three patients, the optimism bias can be large. Please re-run the evaluation with a clean separation between model selection and evaluation (e.g., a separate validation patient or nested cross-validation), or explicitly relabel the findings as exploratory and remove the 'accurately detect' claim from the abstract and conclusion.","section":"Section 3.2 and Section 4"},{"comment":"The evidence for the central accuracy claim is weak even apart from the selection issue. The average DSC is 0.59 ± 0.19, only three of six scans exceed 0.7, and the authors note that Case 2 had poor DSC due to pulp over-segmentation, but no per-scan Dice values are reported. Without the per-case distribution, readers cannot determine whether the average reflects consistent performance or a few favorable folds. Please report per-scan (per-timepoint) DSC, precision, and recall, and temper the conclusion to match the observed variability.","section":"Section 4, Table 1"},{"comment":"The post-processing uses binary erosion and dilation with a fixed kernel size of 6 voxels, but the manuscript gives no justification for this value and no sensitivity analysis. If this parameter was also chosen based on qualitative results on the same six scans, it constitutes another source of test-set leakage. Please justify the choice or evaluate its sensitivity to the reported results.","section":"Section 3.3"}],"minor_comments":[{"comment":"The abstract says '6 longitudinal CBCT datasets,' but the study involves 3 patients with 2 time-points each; please clarify to avoid implying 6 independent patients.","section":"Abstract"},{"comment":"The phrase 'Gaussian smoothing (kernel size = 1 voxel)' is ambiguous; please specify whether this is the kernel radius, the full window size, or the standard deviation of the Gaussian.","section":"Section 3.2"},{"comment":"The text says precision and recall are 'weighted by the class support and averaged over all six cases.' Please clarify how the two time-points of the same patient are handled in leave-one-out cross-validation, and confirm that both time-points of a training patient are in the same fold.","section":"Section 4"},{"comment":"The caption says the scans shown for Case 1 and Case 2 achieved the best and worst Dice, but the main text mentions three scans with DSC > 0.7; please clarify which specific time-points are shown and how 'best' and 'worst' are defined.","section":"Figure 2 caption"},{"comment":"Please provide a specific version or citation for the DentalSegmentator extension to improve reproducibility.","section":"Section 3.1"},{"comment":"There is a typo in 'due to is widespread availability'; it should be 'its widespread availability.'","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is exactly the load-bearing issue: the paper tunes both the feature set and the neighborhood radius on the test data, and the three-patient sample amplifies the risk. I would not reject the manuscript outright because the clinical problem is important and the methodology is clearly described, but the evaluation must be redone with a clean protocol or the claims must be substantially weakened. The paper would also benefit from reporting per-case metrics and from a more cautious framing as a preliminary feasibility study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a proof-of-concept for a real clinical need—automating ECR lesion segmentation in CBCT—using textbook texture features (GLCM/GLRLM) plus a linear SVM. The application is new, the writing is clear, and they use open-source tools, which is a plus. But the evaluation is compromised by a protocol that lets the test data influence the choices that produce the reported number. Section 3.2 selects LGRE/HGRE by eyeballing histograms over the whole dataset, and Section 4 picks the neighborhood radius that gives the best leave-one-out Dice on the same folds. That is test-set tuning. With three patients, the reported average Dice of 0.59 is not a trustworthy estimate of performance on new patients, and the abstract's 'accurately detect' claim is stronger than the evidence supports.\n\nWhat is actually useful: the pipeline is simple and reproducible, the idea of using run-length emphasis features to catch subtle signal changes in ECR is sensible, and the longitudinal clustering of intra-lesion features is a nice exploratory step toward prognostic biomarkers. The authors acknowledge the cohort is small but do not flag the tuning, which is the more serious problem. Other soft spots: no comparison to any baseline (even a simple intensity-threshold), one of six cases fails, and the post-processing (erosion/dilation plus largest connected component) is applied without sensitivity analysis.\n\nThe fix is straightforward: split data by patients before any feature selection or radius choice, or use nested cross-validation, and report results with that separation. If they redo it cleanly, the approach may well work—the features are standard and the clinical motivation is genuine. As is, the paper is a promising preliminary report, not a validated method.\n\nFor a journal: I'd send it to peer review because the problem is worth solving and the flaw is fixable. A reviewer should ask for a proper model-selection protocol and more data. I wouldn't cite it in my own work until that re-evaluation is done.","headline":"A sensible proof-of-concept spoiled by test-set tuning; the clinical problem is real and the fix is straightforward.","tokens_in":4879,"tokens_out":1728,"would_cite":false,"duration_ms":16176,"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 shows that two local texture features, LGRE and HGRE, can be used with a linear support vector machine to segment external cervical resorption lesions in CBCT scans, reaching an average Dice score of 0.59 in leave-one-out…","keywords":["external cervical resorption","cone-beam computed tomography","texture features","grey-level run length matrix","support vector machine","lesion segmentation","k-means clustering"],"falsifier":"Take a new patient with ECR whose CBCT scan was never used to select features or parameters, run the exactly specified pipeline (LGRE and HGRE features, five-voxel radius, linear SVM trained on the three study patients), and compare the predicted segmentation to manual ground truth: if the average Dice score falls to near zero or chance overlap, the claim that texture features accurately detect ECR would be refuted, while a held-out Dice above 0.7 would support it.","tokens_in":3958,"feed_emoji":"🦷","tokens_out":5966,"duration_ms":57110,"temperature":0.7,"pith_summary":"The paper is trying to establish that automated segmentation of external cervical resorption (ECR), a tooth-destroying resorptive process, is possible from cone-beam CT using local texture statistics rather than manual tracing or deep learning. The authors extract two gray-level run-length features, low grey-level run emphasis and high grey-level run emphasis, for each voxel, then train a support vector machine on patient scans and test it on six longitudinal scans with leave-one-out evaluation, reporting an average Dice score of 0.59 with three of six scans above 0.7. They also show that k-means clustering of these texture features within one lesion separates a likely calcified region from radiolucent areas, suggesting the same features could help identify healing tissue. If these claims hold, clinicians could monitor ECR progression automatically and work toward a biomarker that predicts whether resorption will heal or advance.","feed_headline":"Two texture features automate ECR lesion segmentation in CBCT","feed_subtitle":"A linear support vector machine on gray-level run maps segments lesions, with three of six scans scoring above 0.7 Dice.","key_machinery":"The load-bearing objects are the low grey-level run emphasis (LGRE) and high grey-level run emphasis (HGRE) voxel-wise feature maps, both derived from the gray-level run-length matrix. A gray-level run is a set of consecutive collinear voxels with the same gray level; LGRE emphasizes long runs of dark voxels and HGRE emphasizes long runs of bright voxels. The pipeline computes these maps at a neighborhood radius of five voxels after clipping intensities and normalizing, standardizes the features, trains a linear SVM using leave-one-out cross-validation, then erodes, keeps the largest connected component, and dilates to remove over-segmentation. The same feature maps are later clustered with k-means into two groups to separate likely calcified tissue from radiolucent tissue.","core_discovery":"The central claim is that certain texture features can be used to accurately detect subtle CBCT signal changes due to ECR. Concretely, the paper reports that the LGRE and HGRE local texture feature maps, classified by a linear SVM and post-processed with erosion, largest-connected-component selection, and dilation, segment ECR lesions in six longitudinal CBCT scans with an average Dice score of 0.59 ± 0.19, with precision and recall around 0.86 and 0.84. In the best cases the predicted segmentation aligns closely with manual ground truth; in the worst case the classifier over-segments to include the tooth pulp, which appears texturally similar to the lesion. The paper further presents a preliminary k-means analysis on one case showing a darker cluster within the lesion consistent with calcified tissue, which the authors frame as a first step toward imaging biomarkers of ECR prognosis.","pith_inferences":["A natural next experiment would be to freeze the exact pipeline (LGRE and HGRE features, five-voxel radius, linear SVM) and test it on an independent, multi-scanner dataset, since the feature and radius choices in the paper were made with all six scans in view.","The same two-feature recipe may apply to other low-contrast dental lesions such as root caries or internal resorption, because those also appear as local run-length texture changes rather than simple intensity drops.","Using the full set of eighteen computed texture features with automated feature selection, rather than the two chosen by histogram inspection, could either improve Dice scores or confirm that LGRE and HGRE are sufficient.","If the k-means clusters correspond to histological zones of calcification, then measuring cluster volume ratios across time-points could become a quantitative progression endpoint in future clinical studies."],"forward_implications":["If a patient's ECR lesion can be segmented automatically from a CBCT scan, clinicians can measure lesion volume changes over time without manual tracing.","Because the method relies on intensity statistics rather than raw voxel values, the same features could transfer across CBCT scanners more easily than intensity-thresholding approaches.","Unsupervised clustering of lesion voxels could provide a non-invasive readout of calcification, helping to distinguish healed or healing lesions from actively resorbing ones.","The reported Dice score of 0.59 with three scans above 0.7 shows that useful segmentations are attainable, while the pulp-confusion case defines the specific failure mode that future improvements would need to correct."],"supporting_citations":[{"why":"Defines the gray-level co-occurrence matrix texture statistics that motivate the local texture feature maps used in this work.","marker":"[10]"},{"why":"Defines gray-level run-length measures, including low and high grey-level run emphasis, which are the two features selected for classification.","marker":"[11]"},{"why":"Supplies the voxel-wise texture feature map implementation used to extract LGRE and HGRE features from the CBCT scans.","marker":"[12]"},{"why":"Provides the dental segmentation tool used to isolate the affected tooth before texture feature extraction.","marker":"[8]"}],"fun_headline_variants":["Texture features automate ECR lesion detection in CBCT","LGRE/HGRE features segment cervical resorption automatically","Automated CBCT segmentation of ECR lesions via texture analysis","Local texture cues enable automated ECR segmentation in CBCT","Binary texture classifier pinpoints ECR lesions in CBCT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result depends on the idea that the texture features and the five-voxel neighborhood radius chosen by looking at all six study scans will also separate lesions from healthy tissue in a new patient's scan, a generalizability that the paper does not test.","fun_headline_variants_meta":{"raw":{"variants":["Texture features automate ECR lesion detection in CBCT","LGRE/HGRE features segment cervical resorption automatically","Automated CBCT segmentation of ECR lesions via texture analysis","Local texture cues enable automated ECR segmentation in CBCT","Binary texture classifier pinpoints ECR lesions in CBCT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2924,"prompt_tokens":950,"completion_tokens":1974,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":1894}},"tokens_in":566,"tokens_out":1974,"duration_ms":14599,"temperature":1.0,"reasoning_tokens":1894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:13:59.194260+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a new patient with ECR whose CBCT scan was never used to select features or parameters, run the exactly specified pipeline (LGRE and HGRE features, five-voxel radius, linear SVM trained on the three study patients), and compare the predicted segmentation to manual ground truth: if the average Dice score falls to near zero or chance overlap, the claim that texture features accurately detect ECR would be refuted, while a held-out Dice above 0.7 would support it.","supporting_citations":[{"cited_title":"Textural features for image classification,","cited_arxiv_id":null,"evidence_quote":"Defines the gray-level co-occurrence matrix texture statistics that motivate the local texture feature maps used in this work."},{"cited_title":"Texture analysis using gray level run lengths,","cited_arxiv_id":null,"evidence_quote":"Defines gray-level run-length measures, including low and high grey-level run emphasis, which are the two features selected for classification."},{"cited_title":"Computing textural feature maps for n- dimensional images,","cited_arxiv_id":null,"evidence_quote":"Supplies the voxel-wise texture feature map implementation used to extract LGRE and HGRE features from the CBCT scans."},{"cited_title":"Dentalsegmentator: robust open source deep learning-based ct and cbct image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the dental segmentation tool used to isolate the affected tooth before texture feature extraction."}],"review_version":1}