Automated external cervical resorption segmentation in cone-beam CT using local texture features

Asma A. Khan; Beatriz Paniagua; Matthew C. Davis; Sadhana Ravikumar

REVIEW 3 major objections 6 minor 12 references

Automated external cervical resorption segmentation in cone-beam CT using local texture features

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict A sensible proof-of-concept spoiled by test-set tuning; the clinical problem is real and the fix is straightforward. read the letter →

arxiv 2501.05236 v1 pith:D4JFHYAV submitted 2025-01-09 cs.CV

Sadhana Ravikumar , Asma A. Khan , Matthew C. Davis , Beatriz Paniagua This is my paper · ORCID

classification cs.CV

keywords externalcervicalresorptioncone-beamcomputedtomographytexturefeaturesgrey-levelrunlengthmatrixsupportvectormachinelesionsegmentationk-meansclustering

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

What 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.

The 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.

For 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.

Referee Report

3 major / 6 minor

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.

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 (3)

[Section 3.2 and Section 4] 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 4, Table 1] 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 3.3] 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.

minor comments (6)

[Abstract] 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 3.2] 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 4] 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.
[Figure 2 caption] 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 3.1] Please provide a specific version or citation for the DentalSegmentator extension to improve reproducibility.
[Introduction] There is a typo in 'due to is widespread availability'; it should be 'its widespread availability.'

Circularity Check

1 steps flagged · score 4.0 of 10

The reported Dice is selected from the same leave-one-out results used to report it, so the central accuracy claim is test-set tuned rather than independently predicted.

fitted input called prediction [Section 4, Table 1 and surrounding text]
"We evaluated our method by performing leave-one-out cross validation at the patient level. In each split, we trained the SVM using data from two patients and evaluated on data from the left out patient. ... Based on these results, we selected the optimal neighborhood radius for feature extraction as 5 voxels."

The neighborhood radius is a model parameter selected by inspecting the leave-one-out Dice scores in Table 1, and the same Table 1 is then presented as the held-out accuracy of the method. The reported headline Dice of 0.59 is therefore not an independent out-of-sample estimate for a pre-specified pipeline: it is the performance of a configuration chosen because it scored well on the same six test folds. Feature selection in Section 3.2 was also performed by visually comparing histograms of healthy versus ECR voxels on the full dataset, so test patients influenced which LGRE and HGRE features were used. This is test-set tuning rather than an equivalence by construction, but it makes the central claim of accurate detection partly fitted to the evaluation data.

full rationale

There is no derivation-style circularity: the paper contains no first-principles equations or uniqueness claims, and no load-bearing self-citation chain. The cited texture-analysis module [12] is a software tool, and the DentalSegmentator/nnU-Net citations are external tools, not self-referential evidence. The main circularity concern is statistical: the optimal neighborhood radius was selected from the same leave-one-out test results (Table 1) used to report accuracy, and the LGRE/HGRE features were selected after inspecting histograms over all patients, including left-out test voxels. Consequently, the average Dice of 0.59 is a test-set-tuned estimate rather than an unbiased prediction for new patients. The paper itself acknowledges the small sample size as a limitation, but does not acknowledge this model-selection leakage. Because the core evaluation is partly fitted to the test folds, a moderate circularity score of 4 is appropriate; the SVM training on two patients and testing on the third still provides some independent content, so this is not a fully forced or definitional circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the choice of texture features, several hand-set preprocessing and post-processing parameters, and the assumption that manual ground truth and the three patients are representative. No new physical or biological entities are introduced.

free parameters (6)

Neighborhood radius for texture feature extraction = 5 voxels
Selected based on leave-one-out test performance (Section 4): radii 2, 5, 7, 9 were compared and 5 chosen.
Gaussian smoothing kernel size = 1 voxel
Chosen by hand in preprocessing (Section 3.2).
Erosion/dilation kernel size = 6 voxels
Chosen by hand for post-processing to counter over-segmentation (Section 3.3).
Percentile clipping bounds = 5th and 95th percentiles
Chosen by hand in preprocessing to exclude artifacts (Section 3.2).
Feature set (LGRE, HGRE) = LGRE, HGRE
Selected by visually inspecting histograms of all 18 texture features on the dataset and choosing the two with greatest separation (Section 3.2).
k-means cluster count = 2
Chosen for exploring calcified vs. radiolucent clusters (Section 4).

assumptions (4)

domain assumption Texture features computed from GLCM and GLRLM capture tissue changes relevant to ECR in CBCT images.
The whole method rests on this assumption; the paper does not validate it against histology or other imaging (Section 3.2).
domain assumption The manual ECR lesion segmentations by the expert rater are accurate enough to serve as ground truth for training and evaluation.
Ground truth is manually generated by one rater (Section 3.1); no inter-rater reliability is reported.
domain assumption The 3 patients in the study are representative of the broader ECR population.
Only 3 patients (6 scans) are used, and the authors acknowledge the small sample size (Conclusions).
standard math Classical machine learning assumptions for SVM and k-means hold: standardized features, linear separability, and Euclidean distance are appropriate.
Standard implementations from scikit-learn are used (Sections 3.3 and 4).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automated external cervical resorption segmentation in cone-beam CT using local texture features." pith.science (2026). https://pith.science/paper/D4JFHYAV

@misc{pith2026250105236,
  author       = {Pith},
  title        = {Pith review of: Automated external cervical resorption segmentation in cone-beam CT using local texture features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4JFHYAV}},
  note         = {Machine review of arXiv:2501.05236}
}

read the original abstract

External cervical resorption (ECR) is a resorptive process affecting teeth. While in some patients, active resorption ceases and gets replaced by osseous tissue, in other cases, the resorption progresses and ultimately results in tooth loss. For proper ECR assessment, cone-beam computed tomography (CBCT) is the recommended imaging modality, enabling a 3-D characterization of these lesions. While it is possible to manually identify and measure ECR resorption in CBCT scans, this process can be time intensive and highly subject to human error. Therefore, there is an urgent need to develop an automated method to identify and quantify the severity of ECR resorption using CBCT. Here, we present a method for ECR lesion segmentation that is based on automatic, binary classification of locally extracted voxel-wise texture features. We evaluate our method on 6 longitudinal CBCT datasets and show that certain texture-features can be used to accurately detect subtle CBCT signal changes due to ECR. We also present preliminary analyses clustering texture features within a lesion to stratify the defects and identify patterns indicative of calcification. These methods are important steps in developing prognostic biomarkers to predict whether ECR will continue to progress or cease, ultimately informing treatment decisions.

Figures

Figures reproduced from arXiv: 2501.05236 by the authors.

**Figure 2.** Cross-sectional views of example predictions obtained for each patient overlaid on the corresponding CBCT [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Automated ECR lesion segmentations for Case 1 predicted at two time-points 1 year apart. Unsupervised [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Impact of cone-beam computed tomography on treatment planning for external cervical resorption and a novel axial slice-based classification system,

Goodell, K. B., Mines, P., and Kersten, D. D., “Impact of cone-beam computed tomography on treatment planning for external cervical resorption and a novel axial slice-based classification system,” Journal of endodontics 44(2), 239–244 (2018)

work page 2018
[2]

Invasive cervical resorption,

Heithersay, G. S., “Invasive cervical resorption,” Endodontic topics 7(1), 73–92 (2004)

work page 2004
[3]

Understanding external cervical resorption in vital teeth,

Mavridou, A. M., Hauben, E., Wevers, M., Schepers, E., Bergmans, L., and Lambrechts, P., “Understanding external cervical resorption in vital teeth,” Journal of endodontics42, 1737–1751 (12 2016)

work page 2016
[4]

A review of external cervical resorption,

Chen, Y., Huang, Y., and Deng, X., “A review of external cervical resorption,” Journal of Endodontics47(6), 883–894 (2021)

work page 2021
[5]

A volumetric assessment of external cervical resorption cases and its correlation to classification, treatment planning, and expected prognosis,

Matny, L. E., Ruparel, N. B., Levin, M. D., Noujeim, M., and Diogenes, A., “A volumetric assessment of external cervical resorption cases and its correlation to classification, treatment planning, and expected prognosis,” Journal of endodontics46(8), 1052–1058 (2020)

work page 2020
[6]

The detection and management of root resorption lesions using intraoral radiography and cone beam computed tomography–an in vivo investiga- tion,

Patel, S., Dawood, A., Wilson, R., Horner, K., and Mannocci, F., “The detection and management of root resorption lesions using intraoral radiography and cone beam computed tomography–an in vivo investiga- tion,” International endodontic journal42(9), 831–838 (2009)

work page 2009
[7]

3d slicer: A platform for subject-specific image analysis, visualization, and clinical support,

Kikinis, R., Pieper, S. D., and Vosburgh, K. G., “3d slicer: A platform for subject-specific image analysis, visualization, and clinical support,” Intraoperative Imaging and Image-Guided Therapy, 277–289 (2014)

work page 2014
[8]

Dentalsegmentator: robust open source deep learning-based ct and cbct image segmentation,

Dot, G., Chaurasia, A., Dubois, G., Savoldelli, C., Haghighat, S., Azimian, S., Taramsari, A. R., Sivara- makrishnan, G., Issa, J., Dubey, A., et al., “Dentalsegmentator: robust open source deep learning-based ct and cbct image segmentation,” Journal of Dentistry, 105130 (2024)

work page 2024

Show all 12 references

[9]

nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., and Maier-Hein, K. H., “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods18(2), 203–211 (2021)

2021
[10]

Textural features for image classification,

Haralick, R. M., Shanmugam, K., and Dinstein, I., “Textural features for image classification,” IEEE Transactions on Systems, Man, and Cybernetics3, 610–621 (11 1973)

1973
[11]

Texture analysis using gray level run lengths,

Galloway, M. M., “Texture analysis using gray level run lengths,” Computer Graphics and Image Process- ing 4, 172–179 (6 1975)

1975
[12]

Computing textural feature maps for n- dimensional images,

Vimort, J., McCormick, M., Budin, F., and Paniagua, B., “Computing textural feature maps for n- dimensional images,” Insight Journal , 80014–2 (2017)

2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Impact of cone-beam computed tomography on treatment planning for external cervical resorption and a novel axial slice-based classification system,

Goodell, K. B., Mines, P., and Kersten, D. D., “Impact of cone-beam computed tomography on treatment planning for external cervical resorption and a novel axial slice-based classification system,” Journal of endodontics 44(2), 239–244 (2018)

work page 2018

[2] [2]

Invasive cervical resorption,

Heithersay, G. S., “Invasive cervical resorption,” Endodontic topics 7(1), 73–92 (2004)

work page 2004

[3] [3]

Understanding external cervical resorption in vital teeth,

Mavridou, A. M., Hauben, E., Wevers, M., Schepers, E., Bergmans, L., and Lambrechts, P., “Understanding external cervical resorption in vital teeth,” Journal of endodontics42, 1737–1751 (12 2016)

work page 2016

[4] [4]

A review of external cervical resorption,

Chen, Y., Huang, Y., and Deng, X., “A review of external cervical resorption,” Journal of Endodontics47(6), 883–894 (2021)

work page 2021

[5] [5]

A volumetric assessment of external cervical resorption cases and its correlation to classification, treatment planning, and expected prognosis,

Matny, L. E., Ruparel, N. B., Levin, M. D., Noujeim, M., and Diogenes, A., “A volumetric assessment of external cervical resorption cases and its correlation to classification, treatment planning, and expected prognosis,” Journal of endodontics46(8), 1052–1058 (2020)

work page 2020

[6] [6]

The detection and management of root resorption lesions using intraoral radiography and cone beam computed tomography–an in vivo investiga- tion,

Patel, S., Dawood, A., Wilson, R., Horner, K., and Mannocci, F., “The detection and management of root resorption lesions using intraoral radiography and cone beam computed tomography–an in vivo investiga- tion,” International endodontic journal42(9), 831–838 (2009)

work page 2009

[7] [7]

3d slicer: A platform for subject-specific image analysis, visualization, and clinical support,

Kikinis, R., Pieper, S. D., and Vosburgh, K. G., “3d slicer: A platform for subject-specific image analysis, visualization, and clinical support,” Intraoperative Imaging and Image-Guided Therapy, 277–289 (2014)

work page 2014

[8] [8]

Dentalsegmentator: robust open source deep learning-based ct and cbct image segmentation,

Dot, G., Chaurasia, A., Dubois, G., Savoldelli, C., Haghighat, S., Azimian, S., Taramsari, A. R., Sivara- makrishnan, G., Issa, J., Dubey, A., et al., “Dentalsegmentator: robust open source deep learning-based ct and cbct image segmentation,” Journal of Dentistry, 105130 (2024)

work page 2024