{"id":"49cb6801-870a-4e40-b067-46633e7c2493","arxiv_id":"2607.13010","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Fine-tuning MoGe-2's scale/normal heads on a new synthetic dermoscopy dataset (D-Synth) and small real 3D data gives near-metric scale on SKINL2 (0.87x) and WoundsDB (0.91x), with weaker, partly fitted DDI results (1.95x) on a 14-image test set.","lead":"This paper trains a standard depth-estimation model on synthetic skin images to measure real skin lesions in millimeters from ordinary photos, correcting a 16x scale error down to near 1x for dermoscopic images. It also releases a synthetic dermoscopy dataset with pixel-perfect 3D labels, the most reusable contribution for building clinical measurement tools.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-dataset generalization is unsupported: DDI scale correction only appears after training on DDI pseudo-GT.","rationale":"The reader identified DDI pseudo-GT as the weakest assumption, but the deeper load-bearing issue is that the paper never demonstrates generalization to an unseen dataset. The paper's own ablation (Table 3, row 5 vs. row 6) provides the key evidence: without DDI training data, DDI scale ratio stays at 18.3×, showing the model does not transfer. This concern is independent of pseudo-GT noise; even if ruler segmentation were perfect, the model still fails without DDI fine-tuning. The reader's focus on pseudo-GT error is valid but secondary. My verdict remains CONDITIONAL (UNCHANGED) because the core scale-correction results on SKINL2 and WoundsDB use real depth GT and are internally consistent; however, the paper must add a truly held-out dataset evaluation and temper the generalization language.","tokens_in":8708,"tokens_out":4579,"duration_ms":129618,"concrete_test":"Run a leave-one-dataset-out evaluation, starting with the already-reported Table 3 row 5 configuration: train on D-Synth + SKINL2 + WoundsDB (no DDI) and evaluate on the DDI test split; report the scale ratio (expected 18.3×, per the paper's own data). Extend to a full 3-fold cross-dataset protocol: for each of SKINL2, WoundsDB, DDI, train on D-Synth + the other two real datasets and evaluate on the held-out test split. If any held-out ratio exceeds 3× while the within-dataset fine-tuned ratio is near 1×, the claimed cross-benchmark generalization fails and the method should be reframed as per-dataset scale calibration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract) is that fine-tuning on D-Synth plus a small amount of real clinical data 'generalizes our method across three real-world benchmarks.' However, Table 3 row 5 shows that after training on D-Synth + SKINL2 + WoundsDB (all real depth data except DDI), the DDI scale ratio remains 18.3×, barely reduced from the MoGe-2 baseline of 81.0×. Only when 33 DDI pseudo-GT images are added to training (row 6) does DDI improve to 1.95×. Thus the model does not transfer to an unseen dermatological domain; each benchmark requires its own fine-tuning data. Since all three evaluation datasets contributed training samples, the reported test results measure calibration to each dataset, not generalization to new domains. The scale head is a 3-layer MLP on the CLS token, which likely overfits to dataset-specific cues (e.g., smartphone capture angle, ruler presence) rather than learning a general metric mapping. The limitations section acknowledges the small DDI test set but does not address the absence of a truly held-out dataset. This directly undercuts the claim that single-image metric scale is 'achievable' for dermatology without per-dataset calibration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Carrión and Norouzi introduce DermDepth, a fine-tuning approach for metric-scale monocular 3D reconstruction in dermatology. They construct D-Synth, a 4,000-sample synthetic dermoscopic dataset with rendered metric depth and normal maps. Starting from MoGe-2, they freeze the encoder and train a 3-layer MLP scale head and a ConvStack normal head on D-Synth, then progressively fine-tune the scale head on real data: SKINL2 (263 train), WoundsDB (50), and 33 DDI pseudo-GT samples with ruler-derived scale. On held-out test splits, scale ratios improve from 16.10 to 0.87 (SKINL2), 0.62 to 0.91 (WoundsDB), and the DDI ruler-area ratio from 81.0 to 1.95, while SI-δ1 remains high. The paper also reports normal-map visual improvements, a Fitzpatrick skin-tone disparity reduction (10.90 to 1.02), and lesion width/area/volume estimates broadly consistent with published size ranges.","tokens_in":9074,"tokens_out":7037,"duration_ms":69651,"significance":"The paper's strongest contribution is empirical: it documents that four metric-depth foundation models fail badly on dermatological images (4–16x overestimation on dermoscopy, 0.6–0.75x underestimation on wounds), and that fine-tuning only 0.6% of parameters on a synthetic dataset corrects dermoscopic scale error to 1.11x. The release of D-Synth with pixel-perfect depth and normals, and the systematic evaluation across four baselines, are useful resources for the community. The fairness analysis is a positive addition. However, the significance is tempered by the fact that the reported cross-benchmark results are not truly out-of-distribution: the final model trains on train splits of all three evaluation datasets, and the DDI result requires DDI pseudo-GT labels. The synthetic-only benefit for WoundsDB is negative (0.28x vs 0.62x baseline). If the authors provide a true held-out evaluation or clearly frame the method as per-dataset fine-tuning, the contribution would be solid but more modest.","major_comments":[{"comment":"The generalization claim is stronger than the evidence. The DDI ratio improves to 1.95 only after adding 33 DDI pseudo-GT images to training (Table 3, row 6). With D-Synth plus real depth from SKINL2 and WoundsDB, the DDI ratio remains 18.3x (Table 3, row 5), still far from metric scale. Since the final model was fine-tuned on train splits of all three benchmarks, Table 2 measures per-dataset calibration, not cross-domain generalization. I recommend either (a) reporting a zero-shot result on a dataset not used in training, or (b) explicitly rephrasing the abstract and conclusion to say the method requires per-dataset fine-tuning (including pseudo-GT) rather than 'generalizes across three benchmarks'.","section":"Abstract; Table 3, rows 5–6"},{"comment":"The DDI validation is partially circular. Pseudo-GT scale is derived by segmenting a medical ruler and computing its area; the same ruler-area ratio ρ is used as the evaluation metric. Training on these labels may teach the scale head to exploit ruler-specific cues (e.g., ruler presence, known area 6.6 cm²) rather than producing a generally metric depth. The paper acknowledges segmentation uncertainty in §4 but provides no sensitivity analysis. I ask for an error analysis linking ruler segmentation accuracy to ρ, and ideally an independent check (e.g., known dimensions of other objects in the scene) to validate the DDI scale.","section":"§2.1, §3.1, Table 2"},{"comment":"No confidence intervals or significance tests are reported. The DDI test set has n=14 and WoundsDB n=27; per-Fitzpatrick groups in Table 4 have at most a handful of samples. The disparity reduction (10.90 to 1.02) and scale-ratio improvements could be within noise. Please report bootstrap confidence intervals or per-sample distributions, especially for DDI and the fairness table.","section":"Tables 2–4"},{"comment":"The claim of 'increasing texture richness' is supported only by qualitative visual inspection. No quantitative normal-map error or surface-detail metric is reported on real data; the normal head is frozen during real-data training because ground-truth normals are noisy. Provide a quantitative normal evaluation on D-Synth test or another reliable source, or qualify the claim to say the improvement is visual/qualitative.","section":"§3.5, Figure 3; Abstract"},{"comment":"The lesion-width validation against literature ranges is presented as evidence of metric accuracy, but 4 of 14 predictions fall outside the cited ranges and are rationalized post hoc as 'boundary cases.' No uncertainty estimates are given for the predicted widths. This is a preliminary sanity check, not a quantitative validation of metric accuracy. I suggest reporting error bars and a pre-defined tolerance criterion for agreement.","section":"§3.6, Table 5"}],"minor_comments":[{"comment":"The 'Scale Error Training Trajectory' plot is not described. Please clarify what the plotted curves/histograms represent (median, distribution, or individual samples) and how the training stages map to the curves.","section":"Figure 2 (left)"},{"comment":"SI-AbsRel=0.017 is mentioned once but not defined in the metrics list nor reported for other models. Either define and report it consistently or remove it.","section":"§3.2"},{"comment":"Reference [30] appears to be a generic segmentation paper and does not directly support the claim about difficulty of collecting 3D dermatology datasets. Consider citing a dataset-collection-specific reference.","section":"Reference [30]"},{"comment":"The limitations paragraph acknowledges ruler segmentation uncertainty and the small DDI test set, but it should also state explicitly that the DDI result was obtained after training on DDI pseudo-GT data and therefore does not demonstrate cross-dataset generalization.","section":"§4, Limitations"}],"recommendation":"major_revision","confidential_remarks":"This is a useful empirical study with a clear negative result for foundation models on dermatology and an interesting synthetic data contribution. The main concern is that the paper oversells the generalization claim: the DDI result is essentially fitted via 33 pseudo-GT samples, and all three evaluation datasets contributed training data. With softened claims or a true held-out evaluation, the paper could be acceptable. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful application paper, not a breakthrough. The new D-Synth synthetic dataset with pixel-perfect 3D labels is a real asset, and the SKINL2 result — synth-only fine-tuning dropping scale error from 16.1x to 1.11x on real plenoptic data — is convincing and reproducible. The ablation showing head-only fine-tuning beats full fine-tuning is also clean.\n\nThe soft spots are where the abstract overreaches. Table 3 row 5 shows that after training on D-Synth plus real depth data (SKINL2+WoundsDB), DDI scale ratio is 18.3x — barely better than the 81x baseline. The 1.95x number only appears after adding 33 DDI pseudo-GT images to training. So DermDepth does not generalize to an unseen dermatological domain; it calibrates to each dataset. The paper says \"fine-tuning on a small amount of real clinical samples generalizes our method across three real-world benchmarks,\" but that's only true if you allow each benchmark to contribute training data. A truly held-out dataset would be needed to support the generalization claim, and none is provided.\n\nThe DDI pseudo-GT is also shaky: ruler-area ratios derived from segmentation, n=14 test set, no error bars. The paper acknowledges the small n and segmentation uncertainty in the limitations, so I don't think this is dishonesty, but the abstract's \"measurements broadly consistent\" is doing a lot of work. The normal-map \"texture richness\" is qualitative only.\n\nProportionate verdict: the paper is worth reading and citing for D-Synth and the SKINL2/WoundsDB scale correction, and the fairness analysis is a plus. But the core claim of single-image metric scale without per-dataset calibration is not established. For a revised version, I'd want a held-out dataset evaluation (e.g., train on two, test on the third) and error bars on DDI.\n\nRecommendation: send to peer review. A serious referee can push on the generalization claim; the dataset and synthetic-to-real results deserve scrutiny and the authors have released code and data. Just don't expect the abstract to survive intact.","headline":"Solid dataset-plus-fine-tuning paper with a real scale-correction result, but the cross-dataset generalization claim doesn't survive contact with Table 3: DDI improves only after adding DDI pseudo-GT to training.","tokens_in":9565,"tokens_out":1960,"would_cite":true,"duration_ms":17995,"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":"One ordinary photo of a skin lesion can be turned into metric-scale 3D depth, surface normals, and lesion measurements after training on synthetic dermoscopic data.","keywords":["metric-scale 3D reconstruction","dermatology","monocular depth estimation","synthetic data","skin lesion measurement","surface normals","fairness","dermoscopy"],"falsifier":"Take a new set of at least 50 dermoscopic images with ground-truth metric depth measured by a structured-light or plenoptic scanner (not a ruler), run DermDepth, and compute the median predicted-to-true depth ratio; if it falls outside roughly 0.9×–1.1×, the single-image metric-scale claim is falsified on real data.","tokens_in":8579,"feed_emoji":"📏","tokens_out":7079,"duration_ms":69868,"temperature":0.7,"pith_summary":"This paper aims to prove that dermatological measurement—tracking lesion size, volume, and surface texture over time—does not require 3D scanners or multiple captures. The authors construct D-Synth, a synthetic dermoscopic dataset with pixel-perfect metric depth and surface normals, and fine-tune only the scale and normal output heads of an existing affine-invariant monocular depth model on it, followed by a small amount of real clinical data. They report that dermoscopic scale error drops from over 16× to under 1.1× on real images, geometric quality is preserved, normal-map detail increases, and lesion width estimates on 14 real test lesions mostly fall within published clinical size ranges. The significance, if correct, is that a single ordinary photograph could support lesion monitoring, ABCDE screening, and wound measurement in settings where 3D hardware is unavailable.","feed_headline":"One photo yields metric 3D lesion measurements","feed_subtitle":"A fine-tuned model corrects depth scale from 16× error to under 1.1×—no special hardware needed.","key_machinery":"The load-bearing mechanism is D-Synth, a rendered synthetic dermoscopic dataset providing pixel-perfect metric depth, surface normals, and camera intrinsics for 4,000 lesions across 131 melanin levels and 240 morphologies. DermDepth fine-tunes only the scale-prediction MLP and normal-decoder ConvStack of an existing affine-invariant depth foundation model (0.6% of parameters), using a symmetric log-space scale loss and an angular normal loss. Progressive refinement on real datasets, including 33 DDI samples whose metric scale is anchored by visible medical rulers, extends the correction across imaging distances from millimeters to meters.","core_discovery":"The paper claims that monocular metric-scale 3D reconstruction of skin is achievable by combining a new synthetic dataset with pixel-perfect 3D annotations (D-Synth) and parameter-efficient fine-tuning of a foundation model's scale and normal heads. It shows that all four state-of-the-art metric-depth models tested systematically misestimate scale on dermatological images—overestimating dermoscopic depth by 4–16× and underestimating wound depth—while DermDepth brings scale ratios to 0.87× (dermoscopy), 0.91× (wounds), and 1.95× (ruler-area ratio on DDI), and reduces Fitzpatrick skin-tone fairness disparity from 10.90 to 1.02.","pith_inferences":["The DDI quantitative claims rest on ruler-derived pseudo ground truth; an independent dataset with true 3D scans (e.g., structured light) would be the definitive next validation, and the paper's own limitation statement asks for more clean 3D data.","Because D-Synth is rendered only at 12–20 mm dermoscopic distances, extending the synthetic renderer to macroscopic ranges could make sim-to-real transfer self-sufficient and reduce dependence on difficult-to-collect real 3D data.","The reported normal-map texture improvement is qualitative; a quantitative descriptor (e.g., normal-map roughness or anisotropy) could be tested against clinical features such as ulceration or papillomatous morphology.","The same output-head fine-tuning recipe might transfer to other medical imaging domains that can generate synthetic pixel-perfect depth, such as endoscopy or retinal imaging."],"forward_implications":["If DermDepth is right, dermatologists could record lesion width, area, and volume from a single photograph, enabling longitudinal tracking in primary care and telemedicine without 3D hardware.","The reported scale correction (16.1× error to about 1×) implies domain-specific fine-tuning can adapt metric-depth foundation models to dermatology, while the raw baselines are not clinically usable for measurement.","Improved surface-normal maps could supply quantitative texture and border-morphology features for ABCDE screening, supplementing 2D image analysis.","Fairness results (disparity 10.90 → 1.02) suggest that diverse synthetic skin tones plus a small real-sample fine-tune can reduce skin-tone bias in 3D dermatology models.","With only 2.1M trainable parameters, the model is lightweight enough that metric 3D measurement could run on mobile devices in low-resource settings."],"fun_headline_variants":["Single photo, metric 3D skin lesion measurements","3D skin depth error cut from 16x to under 1.1x","Synthetic dermoscopic data yields accurate 3D from one image","DermDepth: one image, metric-scale 3D for skin","3D skin metrics from a single photo: scale corrected"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The main quantitative claims lean on estimating metric scale by finding and segmenting a medical ruler in 47 real skin images; if that ruler-based scale is inaccurate or those images are not representative, the reported correction to under 1.1× and the fairness numbers do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Single photo, metric 3D skin lesion measurements","3D skin depth error cut from 16x to under 1.1x","Synthetic dermoscopic data yields accurate 3D from one image","DermDepth: one image, metric-scale 3D for skin","3D skin metrics from a single photo: scale corrected"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1578,"prompt_tokens":794,"completion_tokens":784,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":692}},"tokens_in":538,"tokens_out":784,"duration_ms":8207,"temperature":1.0,"reasoning_tokens":692,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:08:21.237669+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a new set of at least 50 dermoscopic images with ground-truth metric depth measured by a structured-light or plenoptic scanner (not a ruler), run DermDepth, and compute the median predicted-to-true depth ratio; if it falls outside roughly 0.9×–1.1×, the single-image metric-scale claim is falsified on real data.","supporting_citations":[],"review_version":2}