List of Papers Browse by Subject Areas Author List
Abstract
Dermatological practice routinely involves measuring and tracking lesion size, morphology and texture, as critical components of wound or skin cancer screening, monitoring and diagnosis. To accomplish this task, practitioners often image the skin surface with commonly available off-the-shelf camera sensors. This has led to an overwhelming research focus on 2D methods while these objectives naturally benefit from 3D information. In this paper, we demonstrate that dense monocular 3D reconstructions, metric scale measurements and rich surface normal texture estimates are achievable for both dermoscopic and macroscopic cases without the need for additional hardware or multiple captures. We present DermDepth, the first single-view metric scale 3D model for the dermatological domain and D-Synth, the first synthetic dermoscopic dataset with pixel-perfect 3D information. Our experiments show training DermDepth on D-Synth corrects metric scale error from over 16× to under 1.1× for real dermoscopic data, while preserving geometric quality and increasing texture richness. Fine-tuning on a small amount of real clinical samples generalizes our method across three real-world benchmarks spanning the few mm to hundred cm range, diverse skin-tones, chronic wound cases and produces measurements broadly consistent with disease size reported in medical literature. All code, data and models are available at https://github.com/hectorcarrion/dermdepth
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5594_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/5594_supp.zip
Link to the Code Repository
https://github.com/hectorcarrion/dermdepth
Link to the Dataset(s)
https://huggingface.co/datasets/hcarrion/D-Synth
BibTex
@InProceedings{CarHéc_DermDepth_MICCAI2026,
author = { Carrión, Héctor AND Norouzi, Narges},
title = { { DermDepth: Toward Monocular Metric Scale 3D Reconstruction Models for Dermatology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors present the DermDepth model that is capable of generating depth and normal maps from monocular views of clinical and dermoscopy images without the need for an additioanl hardware. They devised a method to generate synthetic images and effectively training the model using this synthetic data.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Use of sythetic data to finetune the dermdepth model is novel. This decreases the need for laborious labelled data generation.
Authors compared their method against the state of the method in the literature to highlight the success of their method.
Dermdepth model seems to be capable of generating high quality depth and normal maps.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The authors use MoGe-2 as a base model build dermdepth on that. Technically novelty is limited to the retraining of the MLP layers of the MoGe2 model using synthetic and limited amount of real data.
It is not clear how the volume measurements presented in Fig2(right) are conducted. The lesions in the images does not look very different in size where as the volume measurements are very different.
The presented results look visually good. Visualizations like the ones presented in the anonymousgithub page could have been included in the paper to make the results more compelling
Results on real data are limited to small number of samples.
What is the computational overhead of running DermDepth compared to standard 2D analysis tools currently used in clinical settings? Is real-time inference feasible on mobile devices?
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
lesion size measurements can be made in 3D more precisely leading to better tracking of the lesion changes. The authors’ method to convert monocular images to 3D views is a valuable in this respect. Demonstration of usability of synthetic data to train such a model is important.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper presents DermDepth, a monocular metric-scale 3D reconstruction method for dermatology, and D-Synth, a synthetic dermoscopic dataset with pixel-perfect depth, normal maps, and camera intrinsics. The authors show that recent monocular metric depth foundation models (MoGe-2, MapAnything, DA3, PPD) systematically fail on dermatological images. Building on MoGe-2, they fine-tune only 0.6% of parameters (scale head and normal head) on D-Synth followed by progressive real-data training on SKINL2, WoundsDB, and DDI, recovering metric scale across all three benchmarks and reducing Fitzpatrick disparity from 10.9 to 1.02.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
D-Synth is a practical contribution, extending S-SYNTH with pixel-perfect metric depth, normals, and intrinsics, a training signal that is hard to obtain at scale in this domain.
The fine-tuning strategy is conservative and well-suited to the small real-data regime, and the fairness improvement on DDI across Fitzpatrick groups is a meaningful result.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
No real downstream task is evaluated. The paper is motivated throughout by ABCDE measurement, classification, and wound tracking, and explicitly cites Pereira et al. as evidence that 3D improves classification. A simple experiment comparing RGB vs the generated data on classification or segmentation is missing, so the paper only shows metric scale is recoverable, not that it is clinically useful.
The baseline comparison is not fully fair. All four foundation baselines are run zero-shot while DermDepth is fine-tuned.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper contributes a useful synthetic dataset and demonstrates that a lightweight fine-tune is enough to recover metric scale and reduce skin-tone disparity. My main concern is that the “clinical use case” is limited to a sanity check against wide literature ranges, and there is no downstream evaluation of the generated data through tasks such as classification or segmentation. I would raise my rating if the authors add a quantitative downstream experiment with fair baselines.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper performs a comparison of depth estimation models in the context of skin image analysis. In parallel, the authors generate a synthetic dataset by expanding a previous methodology [19] and use this dataset (D-Synth) to train a depth estimation model specific for dermatology data. Experiments are conducted using three datasets, two of them with depth and scale ground truth, validating the ability of the proposed approach to estimate the depth from a single RGB image. On a second stage the method is also employed to estimate the depth of skin lesions on a non-annotated set, comparing them with the values reported in the literature. The authors provide an anonymous link with the synthetic dataset and the models.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
Depth estimation is a fairly unexplored problem in skin image analysis. Thus, this work addresses a relevant research topic.
-
The authors generate a synthetic depth dataset that will be publicly released.
-
They perform a comparison of depth estimation foundation models against the one trained on the synthetic data, as well as a fine-tuned version of the latter. This comparison is performed across 3 skin datasets (dermatoscopy, wound classification, and DDI).
-
The proposed approach is assessed in a fairness setting, evaluating the performance across different skin tones.
-
There is a final evalution where the proposed approach is used to estimate the depth for a dataset without depth values. In this case, the results are compared against values reported in the lierature, showing a good alignment.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
While the creation of a synthetic depth dataset is promising, the generated data is far from real (see Fig. 1). However, this is never discussed in the paper.
-
It is not clear from the paper what is the exact training pipeline of the model. Do the authors perform a first pre-train on the synthetic data followed by fine-tuning on real data?
-
The maps shown in Fig. 3 are nothing like the ground truth. Shouldn’t this be discussed in the paper os ir this expected?
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses a pressing need in skin image analysis related too the depth estimation in skin images. Few works explore this direction and, together with the release of the synthetic data and models, this paper can be a good contribution. Experiments on different skin tones are also relevant. However, it has some aspects that need to be discussed in further detail:
1) In Fig, 1, the generated images lack realism and clinical plausibility. However, the authors never address this clear issue nor discuss whether if affects the performance of the model.
2) It is not fully clear what is the backbone/architecture that is considered in the dinal DermDepths model. In Section 2.2 it is reported that it follows the MoGe-2 model, but the number of parameters is significantly lower. Thus, the authors should clarify this better.
- The depth maps shown in Fig. 3 are significantly different from the ground truth ones. Neverthless, this is never commented by the authors.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the Reviewers and Meta-Reviewer for their decision to early accept our work and their constructive feedback. We are encouraged the work is described positively, addressing a “pressing need in skin image analysis” (R3), with our strategy “well-suited to the small real-data regime” yielding “meaningful results” (R2), and “generating high quality depth and normal maps” (R1). We now respond to concerns:
A. Realism of our synthetic data: D-Synth’s primary value is pixel-perfect 3D ground truth (metric depth, normals, intrinsics). We agree 3D photorealistic capture in dermatology is valuable but difficult to scale. We adopt all of S-Synth’s realism features, including: anatomically-grounded probabilistic lesion growth, layered melanosome/blood/lipid models in epidermis/dermis/hypodermis, and physics-based light scatter across wavelengths and skin tones. Our progressive real-data fine-tuning strategy closes the sim-to-real gap (Fig. 2): D-Synth supervises scale and geometry; small real-data few-shot training adapts to real morphology. Empirically (Table 2), D-Synth alone reduces SKINL2 scale 16.1x → 1.11x with SI-δ_1 preserved, demonstrating photorealism is not required for scale supervision. We will highlight this in camera-ready.
B. Downstream task use: Sec. 3 and Table 4 validate DermDepth’s predicted lesion diameter against published clinical size ranges across 8 diagnostic categories. Accurate diameter estimation from a single monocular capture is clinically-actionable information (“D” in ABCDE). When 3D information is available, performance increases significantly downstream, for example lesion-classification (Sec. 1, Pereira: +10% acc, +29% sens. with specialized hardware). DermDepth makes 3D estimates accessible to practitioners.
C. Additional baseline training: DA3/MoGe-2/MapAnything/PPD all claim zero-shot metric depth as their published contribution. We test that claim on dermatology and find systematic 4 - 156x scale errors despite high SI-δ_1 (≥92% across baselines: relative geometry is recovered, but the global scale is wrong). DermDepth’s contribution is not to establish which of these methods would be best after fine-tuning, but to close the foundation model natural-image to medical-image scale and texture gap. DermDepth’s design choice as an “adapter” and “output-head” on frozen MoGe-2 (the smallest baseline) is also practical; billion-parameter models require vast amounts of data and compute to train. We answer whether or not practitioners should leverage foundation depth models, while also providing a SOTA solution.
D. Clarifying training pipeline & architecture: Sec 2.2 specifies a 3-stage progressive pipeline: (1) D-Synth → train scale + normal heads; (2) SKINL2 + WoundsDB → scale head only, normal head frozen (real GT normals are noisy); (3) DDI pseudo-GT (ruler-derived) → scale head only. DermDepth = frozen MoGe-2 backbone + 2.1M-parameter scale + normal heads. We will add a pipeline diagram and parameter breakdown in camera-ready.
E. Fig 3 and ground-truth noise: Sec 2.1 documents the GT noise: SKINL2 plenoptic depth has local planar noise; WoundsDB ToF is sparse and spatially offset from RGB. Fig 3’s visual differences between GT and DermDepth should be interpreted with these real-data artifacts in mind. The scale multipliers and metric ranges reported alongside Fig 3 can be reliably computed from the data and accurate scale prediction across dermatology ranges is part of our contribution. We will tie this caveat to Fig 3’s caption.
F. Novelty: Contributions: (1) D-Synth dataset and rendering method; (2) systematic foundation-depth benchmark on dermatology; (3) DermDepth’s scale + normal head training on D-Synth, with real-data fine-tuning, achieves SOTA at accurate scale across dermatological ranges; (4) qualitative and quantitative experiments with skin-tone fairness ablation; (5) open release of all artifacts.
We thank the reviewers again for helping us strengthen DermDepth.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper studies an interesting and less explored problem, estimating 3D information from skin images. The use of synthetic data and the fairness analysis are strong points, and the results are promising.
Some aspects could benefit from further clarification, particularly regarding the evaluation setting, the realism of the synthetic data, and certain implementation details. Overall, the strengths outweigh the limitations, and the paper meets the acceptance bar.