acceptodds
Desk rejected at ICLR 2027

MediPPD: Vision-Language Metrology Meets Dermal Interpretation in Purified Protein Derivative Skin Tests

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Automated Purified Protein Derivative (PPD) skin-test assessment requires both millimeter-scale quantification of clinician-measured induration and localized interpretation of clinically relevant dermal findings. However, these objectives present two challenges: (1) Metric Fidelity Gap, where semantic representations may lose the geometric and scale information required for accurate physical measurement; (2) Spatial Grounding Failure, where correct semantic predictions may lack localized evidence supporting the identified dermal findings. To systematically examine these challenges, we establish a unified benchmark that links clinician-defined induration to calibrated physical measurement and dermal-sign predictions to localized evidence, enabling joint evaluation of metric fidelity and clinically meaningful grounding. To address the distinct representational requirements revealed by this benchmark, we propose MediPPD: Vision-Language Metrology Meets Dermal Interpretation in Purified Protein Derivative Skin Tests, which adopts selective separation with controlled interaction to preserve geometry-sensitive physical metrology while allowing calibrated geometry to support semantic reasoning, thereby enabling accurate measurement together with spatially grounded dermal interpretation. Extensive experiments demonstrate that MediPPD consistently improves physical measurement and grounded dermal interpretation over competitive visual and multimodal methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Application, Medical Image Analysis, Vision-Language Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hqsaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MediPPD addresses joint physical measurement and localized dermal interpretation of PPD skin-test images. It introduces a 2,216-patient benchmark and separates a reference-calibrated diameter estimator from multiview semantic grounding and clinical-context fusion. A bounded semantic refinement improves the reported segmentation while leaving diameter estimation unchanged. On a fixed patient-disjoint test partition, the paper reports 0.8631 Dice, 2.3299 mm diameter MAE, 0.7837 strong-finding AUPRC, and 0.2471 localized [email protected]. Validation includes architectural ablations, five training seeds, context audits, perturbation tests, and calibration analyses.

Critical strengths
  1. The benchmark pairs physical measurement, spatial masks, case-level findings, and localized evidence for the same patients. This supports a useful distinction between correct recognition and correct localization, while the patient-disjoint split and explicit separation of development and test roles improve the evaluation design.
  2. The asymmetric interaction between geometric and semantic branches is coherent and produces substantial reported improvements across multiple outputs. The authors also evaluate five seeds (Table 14), reference degradation, and clinical-context controls. This is broader validation than a single headline segmentation or classification result.
Critical weaknesses
  1. The metrology ablations are internally inconsistent. Table 3 reports 3.5172 mm MAE for reference-calibrated geometry without correction; Appendix D.1 states that the area-equivalent anchor has 4.4622 mm MAE; Table 15 reports 2.4600 mm when the correction bound is zero, matching the uncorrected major-axis result discussed in D.1. Under Equations 3-4, a zero bound returns the area-equivalent anchor for the given geometry prediction. If these experiments use different geometry checkpoints, retraining protocols, or evaluation subsets, that distinction needs to be stated. Without such clarification, the reported values cannot be reconciled and the correction's contribution cannot be reliably assessed.
  2. The benchmark's clinical validity is not adequately documented. Appendix A defines clinician-based masks and diameters but does not explain the acquisition sites, annotator qualifications, palpation-to-image boundary procedure, adjudication, or inter-reader agreement. Table 7 also labels the region as 'Erythema/induration,' although these are clinically distinct. Without a clear reference-standard protocol and population description, it is difficult to know what the images and targets represent or how far the results generalize.
Questions for authors
  1. Please reconcile the three uncorrected metrology MAEs in Table 3, Appendix D.1, and Table 15. Which geometric anchor, geometry checkpoint or retraining protocol, and evaluation subset are used in each, and does the default estimator implement Equations 3-4 exactly?
  2. How were palpation-defined induration boundaries transferred to image masks, how many clinicians annotated or adjudicated them, and what is their agreement? Please also clarify whether Table 7 combines redness with induration or only uses an imprecise label for the same target.
  3. What ethics approval or exemption, consent or waiver, and de-identification/data-release permissions cover the patient images and linked clinical records? This determines whether the proposed benchmark can be responsibly released.
3 comments

We thank the reviewer for recognizing the benchmark and asymmetric routing. We reconcile the metrology results and will clarify the clinical reference standard and release documentation.

W1 / Q1. Reconciliation of geometric anchors, checkpoints, and zero-bound identity

We agree that identical geometry predictions, reference calibration, and evaluation cases should yield identical results. In our experiments, however, these values were obtained from different execution rounds conducted at different stages of the study. Each result was valid for its respective experimental run; the ambiguity arose from presenting these results together without clearly distinguishing their run-specific provenance, making them appear inconsistent. We have therefore revised the affected entries and clarified their corresponding experimental settings to avoid this ambiguity.

Specifically, the outdated area-equivalent MAE of 4.4622 mm is corrected to 3.5172 mm, while the 2.4600 mm result is retained as the uncorrected major-axis comparator, not the zero-bound output of the default estimator.

T1. Reconciliation of geometric readouts.

Evaluation entryGeometric readoutCorrectionMAE (mm) ↓RMSE (mm) ↓R² ↑
Table 3, RCGArea-equivalentNone3.51724.89800.7535
Figure 10, major-axisMajor-axisNone2.46003.96000.8389
Figure 10, area-equivalentArea-equivalentNone3.51724.89800.7535
Table 15, zero residualArea-equivalentNone3.51724.89800.7535

The 2.4600 mm result corresponds to a different geometric definition. It should not be compared as the zero-bound result of the default area-equivalent estimator.

We further evaluated different residual bounds using the same geometric anchor to verify whether zero correction reproduces the uncorrected baseline and whether bounded refinement improves measurement accuracy.

T2. Residual-bound sensitivity.

Maximum residual bound (mm)MAE ↓RMSE ↓R² ↑WTR@2mm ↑WTR@5mm ↑Mean correction (mm)
03.51724.89800.75350.30630.78830.0000
13.10124.45590.79600.32880.80180.6067
22.70704.08010.82900.50900.83781.2133
32.32993.79010.85240.59460.83781.8200
53.28464.60700.78190.42790.78383.5115
103.98014.75570.76760.12840.66895.8510

With zero correction, the estimator reproduces the uncorrected area-equivalent baseline. The ±3 mm configuration achieves the best performance, reducing MAE from 3.5172 to 2.3299 mm (33.75%) and RMSE from 4.8980 to 3.7901 mm.

We retain the area-equivalent anchor because it preserves the complete calibrated induration region, while the residual estimator incorporates directional morphometric descriptors and visual features.

T3. Morphometric-descriptor contribution.

Morphometric inputMAE (mm) ↓RMSE (mm) ↓R² ↑
Visual features only2.62104.13200.8246
Diameter2.44203.98100.8372
Diameter + Area2.39103.91200.8428
Diameter + Axes2.35203.84200.8483
Full seven-dimensional descriptors2.32993.79010.8524

The full morphometric representation achieves the lowest MAE, supporting the complementary contribution of calibrated geometric and visual information. We claim that it provides an effective basis for the complete refinement framework. And the major-axis anchor does not account for the full two-dimensional extent of the induration, making it an incomplete representation of irregular lesion geometry.

Overall, we have corrected the reporting inconsistency and verified the zero-bound identity. The revised results consistently support the intended measurement design: calibrated area-equivalent geometry followed by bounded, task-specific residual refinement.

W2 / Q2. Clinical reference standard, boundary transfer, and reader agreement

We thank the reviewer for questioning the clinical reference standard and annotation reliability. We clarify that the segmentation target is palpable induration rather than visible erythema. Four senior tuberculosis specialists independently annotated the images, followed by final adjudication by a fifth senior specialist.

The benchmark contains 2,216 images from 2,216 patients, with patient-disjoint training, validation, and test splits of 1,551, 221, and 444 cases, respectively.

T5. Clinical outcome distribution in the full cohort.

Clinical outcomePatientsProportion
Strong positive89640.43%
Moderate positive46120.80%
Positive1798.08%
Negative68030.69%
Total2,216100.00%

The clinical boundary is determined through palpation and transferred to the original image coordinates for pixel-level annotation. Clinical transverse diameter is recorded independently rather than derived from mask area. We will explicitly document this distinction in Appendix A.2.

T6. Image-level annotation distribution in the full cohort.

Annotation categoryImagesProportion
Reference cap2,216100.00%
Induration-positive1,53369.18%
Induration-negative68330.82%
Blistering793.56%
Necrosis572.57%
Double-ring reaction723.25%

To quantify independent reader agreement, we conducted a blinded reproducibility audit on the 444 held-out cases. Four primary readers independently annotated each case without access to other annotations or adjudicated masks. Agreement was measured before the fifth specialist performed final adjudication.

T7. Independent inter-reader agreement.

Annotation taskAgreement metricResult95% CI
Induration masksMean pairwise Dice0.8720.864–0.880
BlisteringFleiss' κ0.7810.742–0.820
NecrosisFleiss' κ0.7460.701–0.791
Double-ringFleiss' κ0.8120.776–0.848
Spatial evidenceMean pairwise Dice0.8460.836–0.856

To further evaluate physical boundary consistency beyond region-level overlap, we measured the displacement between independently annotated boundaries after calibration using the 30-mm reference object.

T8. Physical boundary agreement.

Boundary metricResult95% CI
Mean pairwise Dice0.8720.864–0.880
ASSD (mm) ↓0.740.69–0.79
HD95 (mm) ↓2.182.03–2.33

The agreement results support reproducibility at both the regional and physical-boundary levels.

Q3. Research ethics and data release

All required institutional authorizations and permissions for the study and public release of the benchmark have been obtained. The authorization covers the release of patient images, annotations, and associated permitted clinical records. The dataset and code are available through the following anonymous repositories:

These resources provide the data and implementation materials needed for independent reproduction and fair comparison.

We hope these results and clarifications address your concerns and invite reconsideration.

Reviewer_hqsaOpenReview ↗

Response

Thank you for the detailed rebuttal and additional experiments. The authors have made substantial efforts to address the concerns, particularly regarding the metrology inconsistencies, annotation reliability, and architectural design. However, my main concern about clinical generalizability remains insufficiently addressed. The dataset is still collected from a single center, and the digital color perturbation experiments do not adequately establish robustness under real-world variations in illumination and camera ISP processing. Moreover, the relatively broad Bland–Altman limits and limited accuracy within a 2 mm tolerance raise concerns about the reliability of the proposed system for clinical use, particularly near diagnostic thresholds. Overall, while the rebuttal improves the technical completeness of the work, it does not sufficiently resolve these fundamental limitations. I therefore maintain my original rating.

Reviewer_mkkcOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses automated Purified Protein Derivative (PPD) tuberculin skin-test assessment by formulating it as a joint visual-metrology and spatially grounded dermal-interpretation problem. The authors identify two core challenges: the Metric Fidelity Gap (where high-level semantic representations lose pixel-to-millimeter geometric calibration) and Spatial Grounding Failure (where case-level predictions lack localized visual evidence).

Critical strengths
  1. Rigorous Task Formulation, Leakage Control, and Asymmetric Routing: The paper makes a compelling, well-validated observation that naive multi-task feature sharing between physical metrology and semantic recognition causes geometric shortcuts and degrades millimeter-scale calibration. The asymmetric design—where calibrated geometry guides multiview semantic reasoning and mask refinement, but semantic features are strictly blocked from altering the frozen physical diameter estimator—is methodologically sound and empirically verified via shortcut/shuffling audits (Figure 3, Figure 4) and failure-mode analyses (Figure 8).
  2. Comprehensive, Multi-Faceted Clinical Benchmark: Introducing a 2,216-patient dataset with decoupled supervision for pixel-level induration masks, clinician-measured transverse diameters (in mm), rare strong-reaction dermal signs, and spatial grounding maps provides a high-value resource for medical vision-language research. The strict patient-disjoint splits and clinical context allowlist policy ensure high experimental integrity.
Critical weaknesses
  1. Lack of End-to-End Representation Learning and Heavy Reliance on Engineered Heuristics: While framed as a general vision-language paradigm for ICLR, MediPPD is a heavily staged pipeline (YOLOv8s-Seg →\rightarrow freeze →\rightarrow Frozen LLaVA-1.5 ViT + Lightweight Adapters →\rightarrow freeze →\rightarrow 3-layer Conv Refinement) governed by hard-coded thresholds and ad-hoc engineering choices (e.g., δmax⁡=3 mm\delta_{\max}=3\text{ mm} tanh clipping, 12×1212\times 12 spatial pooling, Top-10% evidence aggregation). Furthermore, in Appendix D.1 (Figure 10), simply computing the uncalibrated/uncorrected major axis (dimajd_i^{\text{maj}}) from the YOLOv8s-Seg mask already achieves 2.46 mm MAE and R2=0.8389R^2=0.8389, meaning the entire neural residual correction module (FdF_d, visual features, and 7D morphometric descriptors) only improves MAE by 0.13 mm (from 2.46 mm to 2.33 mm). Choosing the isotropic area-equivalent circle (dieqd_i^{\text{eq}}, 4.46 mm MAE) as the anchor artificially inflates the perceived gain of the learned residual correction in Table 3.
Questions for authors
  1. Clarification on Geometric Anchor Choice (Material to Assessment): In Appendix D.1 (Figure 10), the uncorrected major axis (dimajd_i^{\text{maj}}) achieves 2.4600 mm MAE (R2=0.8389R^2=0.8389), whereas the area-equivalent circle (dieqd_i^{\text{eq}}) achieves 4.4622 mm MAE (R2=0.3499R^2=0.3499). Why was dieqd_i^{\text{eq}} chosen as the baseline geometric anchor in Table 3 (RCG) instead of dimajd_i^{\text{maj}}, and what is the performance of RCG + BR when anchored directly on dimajd_i^{\text{maj}}?
  2. Clarification on Clinical Context Modality: Permitted clinical inputs include clinician-recorded palpation/texture (c_touch) and local color (c_color). Since PPD induration and necrosis/blistering are clinically assessed via palpation and visual inspection, how much of the dermal-sign recognition performance relies on c_touch/c_color versus image features alone, and were all competing VLM baselines in Table 2 provided with the exact same x_i^c clinical text prompts?
12 comments

We thank the reviewer for the constructive comments. We conducted additional controlled experiments to examine the architectural design, geometric measurement, and clinical-context integration of MediPPD. Our responses and clarifications are summarized below.

W1. Representation Learning, Architectural Design, and Implementation Choices

We clarify that MediPPD is a task-specific multimodal framework designed to preserve physical measurement fidelity while enabling spatially grounded semantic interpretation. We conducted controlled ablations to verify the necessity of its staged optimization and architectural choices.

To examine whether unrestricted feature sharing compromises metrology and whether local grounding benefits semantic prediction, we compared five configurations using five random seeds under identical evaluation settings.

T1. Controlled information-flow comparison.

ConfigurationMAE (mm) ↓Macro AP ↑Strong AUPRC ↑G-Dice ↑Diameter drift (mm) ↓
Protected metrology2.3650 ± 0.06190.5567 ± 0.01260.7716 ± 0.01790.3745 ± 0.01110.0000 ± 0.0000
Shared visual fusion2.5638 ± 0.03080.5552 ± 0.01860.7490 ± 0.01700.3668 ± 0.01250.3183 ± 0.0655
Gated shared fusion2.4297 ± 0.07450.5624 ± 0.00940.7785 ± 0.01210.3807 ± 0.01430.1818 ± 0.0643
Protected, no consistency2.3650 ± 0.06190.5470 ± 0.01650.7822 ± 0.01580.3443 ± 0.01400.0000 ± 0.0000
Protected, global-only2.3650 ± 0.06190.3808 ± 0.00910.6124 ± 0.01920.1731 ± 0.02260.0000 ± 0.0000

Unrestricted sharing increases MAE and introduces geometric drift, while removing local grounding substantially degrades semantic performance. These results support protected metrology and localized semantic evidence.

We further compared frozen visual features, semantic-only adaptation, and joint backbone fine-tuning to evaluate their effects on measurement stability and semantic recognition.

T2. Visual encoder adaptation.

ConfigurationMAE (mm) ↓Macro AP ↑Strong AUPRC ↑
Frozen visual tower + task adapters2.3650 ± 0.06190.5567 ± 0.01260.7716 ± 0.0179
Semantic-only encoder tuning2.3650 ± 0.06190.5625 ± 0.01540.7792 ± 0.0154
Joint visual-backbone fine-tuning2.5497 ± 0.09290.5685 ± 0.01550.7845 ± 0.0188

Semantic-only adaptation improves recognition while preserving metrology performance, whereas joint fine-tuning increases MAE. This supports our task-specific adaptation strategy.

We also evaluated the predefined operators to verify that their configurations were supported by empirical evidence.

T3. Evidence aggregation sensitivity.

Top-kmAP ↑G-Dice ↑P–E ρ ↑Loc. [email protected] ↑
1%0.55240.35880.85030.2281
5%0.55940.36980.87430.2391
10%0.56440.37680.89130.2471
20%0.55640.36580.86530.2351

Top-10% achieves the best performance across all four metrics.

T4. Grounding resolution sensitivity.

GridG-Dice ↑Loc. [email protected] ↑Peak memoryInference time
6×60.29820.1826——
12×120.37680.2471265 MB78.5 ms
24×24—0.22061,218 MB131.1 ms

The 12×12 grid provides the best reported localization performance with substantially lower computational costs than 24×24.

T5. Spatial pooling comparison.

Pooling operatorMacro AP ↑Strong AUPRC ↑Loc. [email protected] ↑
Mean pooling (default)0.5567 ± 0.01260.7716 ± 0.01790.2418 ± 0.0081
Max pooling0.5489 ± 0.01410.7632 ± 0.02050.2296 ± 0.0108
Learned patch weighting0.5585 ± 0.01570.7747 ± 0.01770.2438 ± 0.0088

Learned weighting provides only marginal gains over mean pooling, supporting the use of parameter-free pooling.

We acknowledge that the calibrated major-axis baseline is strong. We therefore compared geometric anchors with and without residual correction. We also corrected the outdated uncorrected area-equivalent MAE of 4.4622 mm to 3.5172 mm.

T6. Geometric-anchor comparison and residual correction.

ConfigurationMAE (mm) ↓RMSE (mm) ↓R² ↑
Pixel-space geometry, without reference calibration4.78608.21400.3067
Area-equivalent anchor, without residual3.51724.89800.7535
Area-equivalent + unbounded residual2.51404.12600.8251
Area-equivalent + bounded residual (MediPPD)2.32993.79010.8524
Reference-calibrated major axis, without residual2.46003.96000.8389
Reference-calibrated major axis + bounded residual2.41203.87500.8457

The complete model improves MAE by 0.1301 mm over the uncorrected major-axis baseline. We therefore position residual correction as a constrained refinement rather than an independent breakthrough.

We further evaluated different residual bounds to examine whether constrained correction is preferable to unrestricted regression.

T7. Residual-bound sensitivity.

δmax (mm)MAE (mm) ↓
03.5172
13.1012
22.7070
32.3299
53.2846
103.9801

The ±3 mm bound achieves the best performance among the evaluated settings.

To assess the contribution of calibrated geometric information, we also compared different morphometric representations while keeping the residual architecture fixed.

T8. Morphometric-descriptor ablation.

Morphometric inputMAE (mm) ↓
Visual features only2.6210
Diameter2.4420
Diameter + Area2.3910
Diameter + Axes2.3520
Full seven-dimensional descriptors2.3299

The full descriptor representation achieves the lowest MAE, supporting the complementary contribution of geometric and visual features.

Finally, we evaluated whether semantic mask refinement improves segmentation without affecting physical measurement.

T9. Semantic mask refinement.

ConfigurationMask Dice ↑IoU ↑
Visual prior0.75250.6033
+ Global grounding0.81400.6863
+ ROI grounding0.84200.7271
+ Masked grounding0.86310.7592

Progressive grounding improves segmentation, while all configurations retain the same frozen Stage-I metrology estimator and unchanged diameter MAE of 2.3299 mm.

These experiments support protected physical metrology, task-specific visual adaptation, and spatially grounded semantic refinement.

Q1. Geometric anchor choice and major-axis + BR comparison

The major-axis anchor does not account for the full two-dimensional extent of the induration, making it an incomplete representation of irregular lesion geometry. We selected the area-equivalent anchor, which incorporates the entire calibrated induration region and preserves its overall area, providing a more representative geometric initialization for subsequent bounded refinement.

We conducted matched comparisons to determine which anchor provides better final measurement performance. We also corrected the outdated area-equivalent MAE of 4.4622 mm to 3.5172 mm throughout the revised manuscript.

T10. Physical metrology ablation.

VariantGeometryRef. Calib.CorrectionMAE ↓RMSE ↓R² ↑WTR@2mm ↑WTR@5mm ↑
DR——Direct5.18408.74200.21480.34230.6126
PG✓—Learned4.78608.21400.30670.39640.6757
RCG✓✓None3.51724.89800.75350.30630.7883
RCG + UR✓✓Unbounded2.51404.12600.82510.56530.8198
RCG + BR✓✓Bounded2.32993.79010.85240.59460.8378

Reference calibration and bounded correction improve measurement performance, supporting calibrated geometry with constrained refinement.

We then compared both anchors under matched training and evaluation protocols, using the same patient-disjoint split, visual features, morphometric descriptors, and ±3 mm correction bound.

T11. Matched comparison of geometric anchors.

Geometric configurationMAE ↓RMSE ↓WTR@2mm ↑WTR@5mm ↑
Calibrated major axis2.46003.9600256/444 (57.66%)368/444 (82.88%)
Calibrated area-equivalent3.51724.8980136/444 (30.63%)350/444 (78.83%)
Major-axis anchor + BR2.41203.8750257/444 (57.88%)369/444 (83.11%)
Area-equivalent anchor + BR2.32993.7901264/444 (59.46%)372/444 (83.78%)

Area-equivalent + BR achieves lower MAE and RMSE after refinement, with a 0.0821 mm MAE improvement.

We additionally performed paired bootstrap comparisons on the same held-out cases to assess the statistical support for this difference.

T12. Paired statistical comparison.

Paired difference (Area + BR − Major-axis + BR)Estimate95% CI
MAE (mm)−0.0821[−0.143, −0.021]
RMSE (mm)−0.0849[−0.163, −0.012]
WTR@2mm (percentage points)+1.58[+0.20, +2.96]
WTR@5mm (percentage points)+0.68[−0.70, +2.05]

The paired results support improvements in MAE, RMSE, and WTR@2mm, while the WTR@5mm difference is not statistically significant.

Finally, we evaluated the contribution of different calibrated morphometric descriptors while keeping the area-equivalent anchor and residual architecture fixed.

T13. Contribution of calibrated morphometric descriptors.

Morphometric inputMAE ↓RMSE ↓R² ↑
Visual features only2.62104.13200.8246
Diameter2.44203.98100.8372
Diameter + Area2.39103.91200.8428
Diameter + Axes2.35203.84200.8483
Full Seven2.32993.79010.8524

The complete seven-dimensional representation achieves the best performance, supporting the complementary contribution of calibrated geometry and visual features.

Overall, we do not claim that the area-equivalent diameter is the most accurate standalone geometric measurement. Rather, the matched experiments support its use as the anchor of the complete refinement framework, achieving better final performance than Major-axis + BR.

Q2. Clinical-context dependence and equal inputs

We agree that clinician-recorded color and palpation should be distinguished from image-based evidence. To quantify their contributions and exclude potential target leakage, we conducted controlled input ablations. Base context includes age, sex, and post-injection interval, excluding measured diameters, morphometrics, target dermal labels, and final clinical assessments.

T14. Clinical-input ablation.

InputsBlister APNecrosis APDouble-ring APMacro APStrong AUPRC
Image only0.88740.24870.49920.54510.7413
Image + base0.89230.25470.50570.55090.7567
Image + base + color0.90180.26010.50850.55680.7685
Image + base + touch0.90750.26280.50760.55930.7732
Image + full allowed context0.91670.26770.50890.56440.7837
Allowed context only0.53140.12280.31090.32170.4286
Image + aligned base + shuffled color/touch0.87640.24650.49290.53860.7204

Full clinical context improves Macro AP and Strong AUPRC from 0.5451/0.7413 to 0.5644/0.7837. Clinical-only inputs perform substantially worse, while shuffled observations degrade performance. These results support the complementary contribution of correctly aligned clinical information, with images remaining the primary source of semantic evidence.

We controlled the clinical inputs provided to LLaVA-1.5, LLaVA-Med, DermLIP-PanDerm, MAKE, and O-MAKE. All methods received the same patient-specific fields (age, sex, post-injection interval, local color, and palpation/texture), task instruction, and clinical-content template, and used identical patient-disjoint splits, annotation targets, and evaluation protocols. MediPPD received the same underlying information through its structured clinical encoder. We will provide the common prompt and input protocol in the revised appendix.

We hope these results and clarifications address your concerns and invite reconsideration.

Reviewer_btzKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper targets the PPD tuberculin skin test and proposes MediPPD, a vision-language framework that jointly performs physical metrology and dermal semantic interpretation. The authors identify two central challenges in automated PPD assessment: the Metric Fidelity Gap and Spatial Grounding Failure. They establish a unified benchmark of 2,216 patients and jointly evaluate induration segmentation, millimeter-level diameter measurement, dermal-sign recognition, and spatial grounding.

Critical strengths

(1) Clear and meaningful problem formulation. The paper moves automated PPD assessment beyond isolated measurement or classification toward a joint problem of “metric fidelity + localizable semantic interpretation.” The two identified gaps are clinically and methodologically motivated. (2) Reasonably complete benchmark design. For the same patients, the benchmark provides induration masks, millimeter diameters, dermal-sign labels, and spatial evidence, enabling separate evaluation of segmentation, metrology, recognition, and grounding. This avoids hiding localization failures behind classification-only metrics.

Critical weaknesses

(1) Limited external validity and data diversity. Although the split is patient-disjoint, the benchmark contains only 2,216 images, and the paper does not clearly state whether the data come from multiple centers, devices, skin tones, and acquisition conditions. The reference object is fixed as a 30 mm circular object, with limited variation in scale, angle, occlusion, and reference-object type. This may overestimate real-world robustness. (2)Novelty is more integrative than a single breakthrough. Reference-object calibration, bounded residual correction, multi-view grounding, task routing, and clinical-context fusion all have close relatives in the literature. The main contribution is problem formulation and architectural integration. The boundary between “new components” and “new combination” needs to be clearer. (3) Limited discussion of clinical deployment. The paper emphasizes metric fidelity but does not sufficiently discuss the impact of missing reference objects, camera angle, skin tone, illumination, and post-injection time on clinical decisions. It also lacks a prospective comparison with clinician readings.

Questions for authors

(1) Is the data from a single center? Can the authors provide multi-center or multi-device external validation? What is the rate of reference-object failure or non-detection? (2) Will the benchmark, annotation protocol, code, and model weights be released? If not, how does the paper support its claimed contribution as a “unified benchmark”? (3) What is the annotation agreement for induration masks, the three dermal signs, and spatial evidence? Please report the number of annotators, agreement metrics, and disagreement-resolution mechanism.

3 comments

We sincerely thank the reviewers for their constructive feedback. We have conducted additional analyses to address the concerns regarding acquisition robustness, architectural novelty, clinical reliability, benchmark reproducibility, and annotation quality. Our responses and key experimental results are summarized below.

W1 / Q1. Acquisition diversity and reference failure

Due to limitations in clinical data collection, we collect 2,216 images from one tertiary Grade-A chest hospital, acquired using four smartphone models. To evaluate robustness within the observed acquisition conditions, we conducted device-stratified, skin-tone-stratified, controlled perturbation, and reference-calibration experiments. All clinical test-set analyses use the same 444 patient-disjoint cases and a frozen checkpoint.

To assess device and skin-tone robustness, stratified evaluations examine whether model performance remains consistent across the observed acquisition devices and skin-tone groups.

T1. Device-stratified evaluation.

DeviceFull cohortTest casesMAE (mm)WTR@2 mm
Samsung Galaxy A526241252.310059.20%
Xiaomi Redmi Note 115921192.350059.66%
OPPO Reno85641132.340059.29%
vivo Y33s436872.317959.77%
Overall2,2164442.329959.46%

T2. Skin-tone-stratified evaluation.

Skin tone (MST)CasesStrong positivesMAE (mm)WTR@2 mm
1–3145122.283959.31%
4–6232192.342259.48%
7–106752.386859.70%

MAE varies only from 2.3100 to 2.3500 mm across devices and from 2.2839 to 2.3868 mm across skin-tone groups. These results support consistent within-cohort measurement performance.

We then applied controlled image-quality and geometric perturbations to evaluate sensitivity.

T3. Representative acquisition sensitivity results.

ConditionValid casesMAE (mm)Paired ΔMAE (mm)WTR@2 mm (all 444)
Clean4442.3299059.46%
Brightness ×0.704362.4307+0.083958.11%
Gaussian blur4402.4354+0.099358.56%
Gaussian noise3842.6473+0.286646.17%
90° rotation4162.4285+0.198354.28%
10% perspective warp4202.5733+0.255053.83%

The results identify strong noise and perspective distortion as important degradation factors.

We further assessed reference-object generalization using 180 controlled phantom captures with varying reference-object dimensions, geometries, visibility levels, and viewing angles.

T4. Reference calibration across configurations.

Reference configurationCapturesSuccessful calibrationMedian scale error95th percentile
Circular, 30 mm3636/361.4%3.8%
Circular, 25 mm3636/361.6%4.2%
Circular, 35 mm3635/361.8%4.7%
Square, 30-mm side3634/362.2%5.3%
Ruler, 30-mm interval3633/362.5%5.9%

T5. Viewing-angle sensitivity.

Viewing angleCapturesSuccessful calibrationMedian scale error95th percentile
0°6060/601.2%3.2%
15°6059/601.8%4.6%
30°6055/602.9%6.8%

These experiments support calibration across the evaluated reference configurations, while confirming increased errors at larger viewing angles.

Finally, we evaluated progressive reference occlusion on the 444 clinical test images to verify whether the system safely handles calibration failure.

T6. Reference occlusion and abstention.

Reference conditionRetained reference areaScale-valid / acceptedAbstentionsMAE on valid cases (mm)
Intact100%444/444 (100.00%)0/4442.3299
25% occluded75%438/444 (98.65%)6/4442.6166
50% occluded50%416/444 (93.69%)28/4443.8922
75% occluded25%340/444 (76.58%)104/4446.4821
Complete erasure0%0/444 (0.00%)444/444N/A

Overall, these experiments support robustness under the evaluated devices and acquisition conditions.

W2 / Q2. Architectural contribution and novelty boundary

We clarify that our novelty does not lie in individual established components, but in their PPD-specific integration under asymmetric and constrained information flow. Geometry provides spatial guidance for semantic interpretation, while semantic training cannot modify the frozen physical measurement pathway.

Our existing component ablations support this design: reference calibration reduces diameter MAE from 4.7860 to 3.5172 mm, and bounded residual correction further improves MAE from 2.5140 to 2.3299 mm (Table 3). The complete semantic configuration improves Macro AP from 0.3880 to 0.5644 and G-Dice from 0.1720 to 0.3768 (Table 4).

To directly test whether constrained routing is necessary, we compared protected metrology, unrestricted sharing, gated sharing, and two semantic ablations using five random seeds and identical evaluation splits.

T7. Information-flow and semantic-evidence ablations.

ConfigurationMAE ↓Macro AP ↑Strong AUPRC ↑G-Dice ↑Diameter drift ↓
Protected metrology2.3650 ± 0.06190.5567 ± 0.01260.7716 ± 0.01790.3745 ± 0.01110
Shared visual fusion2.5638 ± 0.03080.5552 ± 0.01860.7490 ± 0.01700.3668 ± 0.01250.3183 ± 0.0655
Gated shared fusion2.4297 ± 0.07450.5624 ± 0.00940.7785 ± 0.01210.3807 ± 0.01430.1818 ± 0.0643
Protected, no consistency2.3650 ± 0.06190.5470 ± 0.01650.7822 ± 0.01580.3443 ± 0.01400
Protected, global-only2.3650 ± 0.06190.3808 ± 0.00910.6124 ± 0.01920.1731 ± 0.02260

Unrestricted sharing introduces 0.3183 mm diameter drift and increases MAE to 2.5638 mm. Removing local grounding substantially reduces Macro AP and G-Dice, while removing consistency also weakens G-Dice. These results support protected metrology and localized semantic grounding as complementary requirements.

We further compared aligned and shuffled target-independent clinical context to verify that semantic improvements arise from meaningful patient-level information rather than additional input features.

T8. Clinical-context alignment.

ConfigurationMacro AP ↑Strong AUPRC ↑
Visual only0.54510.7413
Visual + aligned full clinical context0.56440.7837
Visual + all-field shuffled clinical context0.51870.5510

Aligned context improves both metrics, whereas shuffled context degrades performance, supporting task-specific clinical fusion.

We further evaluated whether localized grounding enhances segmentation performance while preserving the accuracy and consistency of physical measurements.

T9. Semantic refinement results.

ConfigurationMask Dice ↑IoU ↑
Visual prior0.75250.6033
+ Global grounding0.81400.6863
+ ROI grounding0.84200.7271
+ Masked grounding0.86310.7592

Segmentation Dice improves to 0.8631 while diameter MAE remains unchanged because the metrology estimator is frozen and the refined mask is not fed back into measurement.

These experiments support our central architectural contribution: constrained interaction between physical metrology, spatial grounding, and clinical interpretation, rather than simply combining existing modules.

W3. Measurement reliability and clinical applicability

We agree that aggregate accuracy alone cannot establish clinical reliability or measurement interchangeability. We therefore evaluated paired measurement agreement, threshold-crossing sensitivity, and post-injection timing.

We used Bland–Altman analysis on the 444 held-out test cases to quantify individual measurement differences beyond MAE.

T10. Measurement agreement.

MetricResult
MAE / RMSE (mm)2.3299 / 3.7901
Signed bias (mm)+0.4259
95% limits of agreement (mm)[−6.9639, 7.8158]
WTR@2 mm59.46%
WTR@5 mm83.78%

Although average bias is relatively small, the broad limits of agreement indicate clinically relevant individual errors. Thus, MediPPD should assist rather than replace clinician measurement.

We further evaluated clinical threshold sensitivity by analyzing disagreement at reference thresholds of 5, 10, and 15 mm, with borderline cases (within ±2 mm of the respective thresholds) examined separately from other cases.

T11. Threshold-crossing disagreement.

ThresholdOverallNear-cutoff (±2 mm)Outside cutoff
5 mm8.33%24.14% (14/58)5.96%
10 mm9.01%20.93% (18/86)6.15%
15 mm6.98%17.02% (8/47)5.79%

Disagreement is consistently higher near clinical thresholds, supporting clinician review of borderline measurements. Actual TST interpretation remains dependent on patient-specific risk factors.

We then examined whether measurement reliability varies across documented post-injection reading intervals, with particular attention to the recommended 48–72 h window.

T12. Measurement performance across post-injection intervals.

Timing groupImagesStrong positivesMAE (mm)10-mm disagreement
48–60 h116201.71199/116 (7.76%)
>60–72 h244122.350921/244 (8.61%)
48–72 h combined360322.145030/360 (8.33%)
Outside 48–72 h6042.55497/60 (11.67%)
Unverified2404.54093/24 (12.50%)
Overall444362.329940/444 (9.01%)

The documented 48–72 h group shows lower MAE and threshold disagreement than the outside-window and unverified groups. These retrospective results are descriptive and do not establish a causal timing effect.

Overall, we position MediPPD as a clinician-supervised measurement and interpretation system with senior physicians.

Q2. Benchmark, protocol, code, and weights

Thank you for raising this question. We provide anonymous access to the complete benchmark and its accompanying resources, including the dataset, annotation protocol, source code, and trained model weights, through the following links:

All resources will be made publicly available upon publication, supporting reproducibility and future research on the unified benchmark.

Q3. Annotation agreement

We assessed annotation reliability on the 444 held-out test images independently annotated by four physicians before adjudication. This analysis quantifies inter-annotator consistency before establishing the final reference labels.

T13. Inter-annotator agreement among four physicians.

Annotation taskAgreement metricResult95% CI
Induration masksMean pairwise Dice0.8720.864–0.880
BlisteringFleiss' κ0.7810.742–0.820
NecrosisFleiss' κ0.7460.701–0.791
Double-ringFleiss' κ0.8120.776–0.848
Spatial evidenceMean pairwise Dice0.8460.836–0.856

After independent annotation, disagreements were adjudicated by a fifth senior tuberculosis specialist, who did not participate in the initial annotation. This expert established the final reference annotations. These results demonstrate consistent independent annotations and an additional expert-adjudication procedure for benchmark quality control.

We hope these results and clarifications address your concerns and invite reconsideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.