Pith. sign in

REVIEW 4 major objections 5 minor 25 references

LesionGen: A Concept-Guided Diffusion Model for Dermatology Image Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LesionGen claims that concept-rich text prompts let a fine-tuned text-to-image diffusion model generate dermatology images that, added to real data, match or beat real-only classifier accuracy and sharply improve rare-class precision.

desk verdict Abstract overstates synthetic-only parity, but the concept-guided augmentation finding is real and worth a careful look. read the letter →

arxiv 2507.23001 v1 pith:TVI5TUN5 submitted 2025-07-30 eess.IV cs.CV

Jamil Fayyad , Nourhan Bayasi , Ziyang Yu , Homayoun Najjaran This is my paper · ORCID
classification eess.IVcs.CV
keywords dermatologyimagesynthesistext-to-imagediffusionmodelsskinlesionclassificationconcept-guidedcaptionsrare-classperformancesyntheticdataaugmentationlow-rankadaptationclass-balancedprompting
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

This paper attempts to establish that a text-to-image diffusion model conditioned on concept-rich dermatological captions can generate synthetic skin-lesion images good enough to train classifiers, with the strongest gains on rare and worst-case subgroups. The proposed pipeline, LesionGen, builds captions from expert clinical attributes and from vision-language pseudo-reports, then fine-tunes a pretrained diffusion model with LoRA and generates class-balanced images using paraphrased prompts. The authors claim that models trained solely on this synthetic data achieve accuracy comparable to real-data training while improving worst-case subgroup performance. In the experiments, the clearest support for the method comes when synthetic images are added to real training data, where overall accuracy and rare-class precision rise above real-only baselines.

What carries the argument

The central object is the concept-guided caption: a textual description organized around seven clinically meaningful attributes, such as pigmentation, lesion elevation, and structure, that define a dermoscopy image. Expert-annotated metadata from D7P is converted into dermatologist-style text through a templated language-model prompt, while HAM10000 images receive pseudo-captions generated by a vision-language model constrained to describe the same seven concepts. These image-caption pairs fine-tune a pretrained latent text-to-image diffusion model with low-rank adaptation (LoRA), using CLIP text embeddings for conditioning; at generation time, prompts are paraphrased to balance class representation. The captions carry clinical information into the generated pixels, and the balancing counteracts the class imbalance that otherwise makes minority-class precision collapse.

What would settle it

Run LesionGen with a strict procedure that splits both datasets into captioning/training and test sets before any caption or fine-tuning step, then check whether synthetic-plus-real still beats real-only overall and on rare classes; if the gains shrink to noise, the reported improvements depend on test-set leakage rather than on the synthetic images themselves.

Watch

Extended reading notes

Core claim

The paper's central claim is that conditioning a fine-tuned diffusion model on structured, concept-rich dermatological captions yields synthetic dermoscopy images that are more useful for downstream classification than images from label-only prompts. In their main experiments, a classifier trained on real images plus LesionGen-generated images reaches 65.3% overall accuracy on D7P versus 65.0% for real-only training, and 75.6% on HAM10000 versus 73.7%, while dermatofibroma precision rises from 0.000 to 0.500 on D7P and from 0.000 to 0.418 on HAM10000. The paper also reports that synthetic-only training reaches only 55.1% and 52.4% on the two test sets, below the real-only numbers, so the direct evidence supports combining synthetic and real data rather than relying on synthetic data alone. The authors frame the mechanism as clinical concepts plus class-balanced prompt paraphrasing, with the intended benefit being worst-case subgroup performance.

Load-bearing premise

The evaluation assumes that no image used for the final accuracy numbers was seen, directly or indirectly, when the captions were written or the diffusion model was fine-tuned, since that split is not stated and the two benchmark datasets draw on overlapping dermatology collections.

Editorial extensions

If this is right

  • LesionGen synthetic images can be used alongside real data to raise overall classification accuracy and, more markedly, precision on rare classes such as dermatofibroma and actinic keratosis.
  • Concept-rich prompts are load-bearing: replacing them with label-only prompts drops synthetic-only accuracy to 32.4% on D7P and 17.6% on HAM10000, so the captioning step is essential to the method.
  • Class balancing through prompt paraphrasing matters: removing it makes minority-class precision collapse even when real data is present.
  • Synthetic data alone is not yet a substitute for real data; the paper's Ablation A shows synthetic-only training underperforming real-only training on both benchmarks.
  • A classifier trained with this augmentation can be more balanced across diagnostic classes, accepting small drops on some majority classes such as vascular lesions in exchange for improved worst-case performance.

Reading between the lines

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

  • If the caption-generation step were moved to a strict train-only split, the measured gains could shrink; the paper does not report whether test images were excluded before the vision-language model wrote captions or before LoRA fine-tuning, so the cleanest next experiment is to caption before any test split and re-run the comparison.
  • The same concept-captioning-plus-balancing recipe could transfer to other medical imaging domains where expert annotations exist for a subset but not for the target dataset, such as histopathology or retinal imaging.
  • Because the generated images are conditioned on clinically structured text, the method provides a natural interface for specifying desired lesion attributes at generation time, pointing toward interactive or clinician-in-the-loop synthesis that the paper lists as future work.
  • If the gains persist under a strict split, the approach suggests a practical answer to data scarcity: use expert-annotated subsets to write rich captions for larger, sparsely annotated collections, then generate balanced training data for rare conditions.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes LesionGen, a text-to-image diffusion framework for dermatology image synthesis. Image-caption pairs are constructed from expert-annotated D7P metadata (via LLM-generated descriptions) and from HAM10000 images (via VLM-generated concept-guided pseudo-captions), and Stable Diffusion v1.4 is fine-tuned with LoRA on these pairs. The resulting model generates class-balanced synthetic skin lesion images, which are then used together with real images to train a ResNet18 classifier. The paper's abstract claims that models trained solely on LesionGen synthetic data achieve classification accuracy comparable to real-image training, with notable gains in worst-case subgroup performance. The main experimental results, however, show the strongest gains in the synth+real setting rather than the synthetic-only setting, and the paper's own Ablation A shows that synthetic-only training underperforms real-only training by 9.9 percentage points on D7P and 21.3 percentage points on HAM.

Significance. If the paper's headline claim were supported, LesionGen would offer a practical route to training dermatology classifiers with reduced reliance on scarce real data, which would be a meaningful contribution to medical image synthesis and classification. The paper also contributes a concrete pipeline for converting sparse structured metadata and unlabeled images into concept-rich captions for diffusion model conditioning, and it makes code and data availability claims. The reported improvement in rare-class precision (e.g., df precision rising from 0.000 to 0.500 on D7P and from 0.000 to 0.418 on HAM in the synth+real setting) is potentially valuable. However, the abstract's synthetic-only parity claim is directly contradicted by the paper's own ablation results, and several evaluation details (data-split timing, baseline sample matching, and run-to-run variability) are unspecified. These issues must be resolved before the central claim can be accepted.

major comments (4)
  1. [Abstract and Section 3.5, Tables 3 and 4] The abstract states that 'models trained solely on our synthetic dataset achieve classification accuracy comparable to those trained on real images,' but Ablation A in Section 3.5 directly contradicts this. Synthetic-only LesionGen-R&B achieves 55.1% accuracy on D7P versus 65.0% for real-only training (Table 3), and 52.4% versus 73.7% on HAM (Table 4), deficits of 9.9 and 21.3 percentage points. Moreover, the highlighted worst-class gains appear only in the synth+real setting: df precision is 0.500 on D7P and 0.418 on HAM in Tables 1 and 2, but 0.000 and 0.045 in the synthetic-only rows of Tables 3 and 4. The abstract and the introduction should be revised to state the defensible claim: concept-guided synthetic data combined with real data improves overall accuracy and rare-class precision, while synthetic data alone is not sufficient, as the paper's own text acknowledges.
  2. [Section 2.1 and Section 3.1] The split between the data used for caption generation and diffusion fine-tuning and the data used for downstream classification evaluation is not specified. Section 2.1 says image-caption pairs are constructed from the full D7P and HAM datasets, and Section 3.1 says only that evaluation is on 'held-out real test sets,' without stating that test images were excluded before VLM captioning or LoRA fine-tuning. If test images entered the caption-generation or diffusion fine-tuning stage, the generative model could memorize or approximate test images and inflate the reported classification accuracy. The authors must specify the exact protocol (e.g., a strict train/validation/test partition performed before caption generation, with test images never used in any generative training step) or re-run the experiments under such a partition. The overlap between D7P and HAM makes this omission particularly consequential.
  3. [Tables 1-4] All classification results are reported as single point estimates with no error bars, standard deviations, or number of random seeds. This is especially problematic for the D7P comparison where LesionGen-R&B (synth+real) achieves 65.3% versus 65.0% for real-only, a difference of 0.3 percentage points that is within expected run-to-run variability for a from-scratch ResNet18. The authors should report mean and standard deviation over at least three independent training runs (or otherwise justify that the differences are stable), for both overall accuracy and per-class precision, before claiming that LesionGen outperforms the real-only baseline.
  4. [Section 3.4 and Tables 1-2] The real-only baseline row is not described with sufficient detail to ensure a sample-matched comparison. Section 3.4 states that the combined (synth+real) setting uses 'up to 250 real images per class (when available), supplemented with synthetic images... to reach a total of 500 training samples per class,' but the real-only baseline in Tables 1 and 2 does not state how many real images per class were used. If the real-only baseline uses more real images (or the full dataset) while the synth+real setting is capped at 250 real images per class, the comparison is unfair and the reported gains could reflect dataset size rather than synthetic data quality. The authors should state the exact training-set composition for every row in Tables 1 and 2.
minor comments (5)
  1. [Section 3.3] The implementation details state that LoRA is used with rank 64, which is unusually high for a low-rank adaptation and approaches full fine-tuning in parameter count; the authors should clarify whether this is intentional and report the number of trainable parameters.
  2. [Section 3.2] The p-SOTA baseline is described as fine-tuning and sampling with label-only prompts, but the number of generated samples per class and the real-image subset used in its synth+real evaluation are not specified; this should be reported to ensure the baseline is comparable to LesionGen-R&B.
  3. [Section 3.6] Figure 3 shows only qualitative examples; since the paper argues that LesionGen's synthetic images carry clinically relevant features beyond human perception, quantitative generation-quality metrics (e.g., FID, KID, or a human/clinical reader study) would strengthen this claim, though the current downstream classification results are the primary evidence.
  4. [Abstract and Conclusion] The phrase 'Code and data are available here' appears in the abstract with the word 'here' as a placeholder; the authors should replace it with a working link or repository identifier.
  5. [Throughout] The paper uses 'significant' and 'significantly' in several places (e.g., Section 3.4 and the Conclusion) without statistical significance testing; given the lack of error bars, these words should be replaced with descriptive terms such as 'larger' or 'improved' unless formal significance tests are reported.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity; the paper's headline claim is internally contradicted by its own Ablation A, but the evaluation is not circular by construction.

full rationale

LesionGen's derivation chain is empirical rather than algebraic: captions are generated from D7P metadata and VLM pseudo-descriptions, a Stable Diffusion model is LoRA-fine-tuned on the image-caption pairs, synthetic images are sampled, and a ResNet18 is trained on real/synthetic mixtures and tested on real labels. No fitted parameter is renamed as a prediction, no target quantity appears in the conditioning or loss, and no load-bearing uniqueness theorem is imported from prior work by the authors. The self-citations in the references are background citations for classification and OOD detection; they do not carry the central argument. The evaluation is against held-out real test labels, and the classifier is not optimized on those labels. Two non-circular concerns remain. First, the abstract's claim that 'models trained solely on our synthetic dataset achieve classification accuracy comparable to those trained on real images' contradicts the paper's own Ablation A results (55.1% vs 65.0% on D7P; 52.4% vs 73.7% on HAM); this is an internal inconsistency, not a circularity. Second, the paper never states that the test images were excluded from the caption-generation or LoRA fine-tuning stage: Section 2.1 builds image-caption pairs from all D7P and HAM images, and Section 3.1 only says evaluation is on 'held-out real test sets'. This is a potential data-leakage risk that would undermine the reported accuracy, but it is not a definitional equivalence between input and output. Because the central comparison is measured rather than constructed, the circularity score is low.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim does not introduce new physical or conceptual entities. It rests on unvalidated assumptions about caption quality, dataset split hygiene, memorization, and single-run evaluation, plus six hand-chosen hyperparameters that are not tested for sensitivity.

free parameters (6)
  • LoRA rank = 64
    Chosen for efficient fine-tuning; no sensitivity analysis reported.
  • Learning rate = 1e-5
    Hand-selected with a constant scheduler; no grid search reported.
  • Diffusion training steps = 15,000
    Hand-selected stopping point based on qualitative validation prompts.
  • Prompt generation temperature = 0.3
    Used for GPT-4o caption generation and paraphrasing; no ablation.
  • Target synthetic samples per class = 500
    Chosen for class balancing; no analysis of sensitivity to this number.
  • Output resolution = 256x256
    Fixed by the generation config; downstream classifier resizes to 224x224.
assumptions (5)
  • domain assumption The seven D7P concepts transfer to HAM10000 and are sufficient for clinically meaningful description.
    Section 2.1 builds both expert and VLM captions on these seven concepts; no dermatologist validation of VLM captions is provided.
  • domain assumption GPT-4o vision-language captions are clinically accurate without expert verification.
    Section 2.1 calls the HAM captions 'pseudo-generated' and does not sample-check them.
  • domain assumption D7P and HAM images used for captioning and diffusion fine-tuning are disjoint from the classification test sets.
    Section 3.1 refers to held-out test sets but never defines the split relative to Section 2.1 data construction.
  • domain assumption Generated synthetic images are not memorized copies of real training or test images.
    Section 3.6 shows only qualitative samples; no similarity or duplication check is reported.
  • domain assumption Single-run ResNet18 training is representative of classifier performance.
    Section 3.3-3.4 report one run per configuration with no seeds or confidence intervals.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LesionGen: A Concept-Guided Diffusion Model for Dermatology Image Synthesis." pith.science (2026). https://pith.science/paper/TVI5TUN5

@misc{pith2026250723001,
  author       = {Pith},
  title        = {Pith review of: LesionGen: A Concept-Guided Diffusion Model for Dermatology Image Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TVI5TUN5}},
  note         = {Machine review of arXiv:2507.23001}
}
read the original abstract

Deep learning models for skin disease classification require large, diverse, and well-annotated datasets. However, such resources are often limited due to privacy concerns, high annotation costs, and insufficient demographic representation. While text-to-image diffusion probabilistic models (T2I-DPMs) offer promise for medical data synthesis, their use in dermatology remains underexplored, largely due to the scarcity of rich textual descriptions in existing skin image datasets. In this work, we introduce LesionGen, a clinically informed T2I-DPM framework for dermatology image synthesis. Unlike prior methods that rely on simplistic disease labels, LesionGen is trained on structured, concept-rich dermatological captions derived from expert annotations and pseudo-generated, concept-guided reports. By fine-tuning a pretrained diffusion model on these high-quality image-caption pairs, we enable the generation of realistic and diverse skin lesion images conditioned on meaningful dermatological descriptions. Our results demonstrate that models trained solely on our synthetic dataset achieve classification accuracy comparable to those trained on real images, with notable gains in worst-case subgroup performance. Code and data are available here.

Figures

Figures reproduced from arXiv: 2507.23001 by the authors.

Figure 1
Figure 1. Overview of LesionGen. Lesion images are paired with either expert derma￾tological descriptions transformed into text by an LLM, or pseudo-captions generated from image features using a VLM, forming a unified multimodal dataset. A LoRA￾tuned diffusion model is trained on the multimodal dataset, and the trained model generates diverse, class-consistent images, which are then evaluated in the downstream classification… view at source ↗
Figure 2
Figure 2. Prompt templates and example outputs for dermatological caption generation from (top) expert-annotated D7P data and (bottom) pseudo-generated HAM data. VLM to produce free-form clinical descriptions, we craft targeted prompts that explicitly instruct the model to describe each image using the seven diagnostic concepts defined in the D7P dataset (e.g., pigmentation, elevation, structure). Although these attributes ar… view at source ↗
Figure 3
Figure 3. Examples of synthetic skin lesion images generated from LesionGen-R&B (top) and p-SOTA (bottom) across six classes. performs well on majority classes; e.g., on HAM, precision for nv reaches 93.8% in the synth+real setting, and vasc reaches 100% even with synthetic data alone. However, performance on minority classes could collapse entirely; e.g., df. Ablation C: Impact of Using Static, Label-Only Prompts. Finally, w… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages

  1. [1]

    In: International Conference on Medical Image Computing and Computer- Assisted Intervention (MICCAI)

    Akrout, M., Gyepesi, B., Holló, P., Poór, A., Kincső, B., Solis, S., Cirone, K., Kawahara, J., Slade, D., Abid, L., et al.: Diffusion-based data augmentation for skin disease classification: Impact across original medical datasets to fully synthetic images. In: International Conference on Medical Image Computing and Computer- Assisted Intervention (MICCAI...

    work page 2023

  2. [2]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)

    Bayasi, N., Du, S., Hamarneh, G., Garbi, R.: Continual-gen: Continual group en- sembling for domain-agnostic skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 3–13 (2023)

    work page 2023

  3. [3]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)

    Bayasi, N., Fayyad, J., Bissoto, A., Hamarneh, G., Garbi, R.: Biaspruner: Debiased continual learning for medical image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 90–101. Springer (2024)

    work page 2024

  4. [4]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)

    Bayasi, N., Hamarneh, G., Garbi, R.: Culprit-prune-net: Efficient continual se- quential multi-domain learning with application to skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 165–175. Springer (2021)

    work page 2021

  5. [5]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Bayasi, N., Hamarneh, G., Garbi, R.: Continual-zoo: Leveraging zoo models for continual classification of medical images. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 4128–4138 (2024)

    work page 2024

  6. [6]

    arXiv preprint arXiv:2411.10004 (2024)

    Chen, R., Zhang, W., Liu, B., Chen, X., Xu, P., Liu, S., He, M., Shi, D.: Eyediff: text-to-image diffusion model improves rare eye disease diagnosis. arXiv preprint arXiv:2411.10004 (2024)

    arXiv 2024

  7. [7]

    Medical Imaging with Deep Learning (MIDL) (2023)

    De Wilde, B., Saha, A., de Rooij, M., Huisman, H., Litjens, G.: Medical diffusion on a budget: textual inversion for medical image generation. Medical Imaging with Deep Learning (MIDL) (2023)

    work page 2023

  8. [8]

    Foundation Models as Class-Incremental Learners for Dermatological Image Classification

    Elkhayat, M., Mahmoud, M., Fayyad, J., Bayasi, N.: Foundation models as class-incremental learners for dermatological image classification. arXiv preprint arXiv:2507.14050 (2025)

    work page Pith review arXiv 2025

Show all 25 references
  1. [9]

    Fayyad, J.: Out-of-distribution detection using inter-level features of deep neural networks. Ph.D. thesis, University of British Columbia (2023)

    2023

  2. [10]

    Computer Methods and Programs in Biomedicine 253, 108231 (2024)

    Fayyad, J., Alijani, S., Najjaran, H.: Empirical validation of conformal prediction for trustworthy skin lesions classification. Computer Methods and Programs in Biomedicine 253, 108231 (2024)

    2024

  3. [11]

    Image and Vision Computing 142, 104897 (2024)

    Fayyad, J., Gupta, K., Mahdian, N., Gruyer, D., Najjaran, H.: Exploiting classi- fier inter-level features for efficient out-of-distribution detection. Image and Vision Computing 142, 104897 (2024)

    2024

  4. [12]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  5. [13]

    Advances in neural information processing systems (NeurIPS)33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems (NeurIPS)33, 6840–6851 (2020)

    2020

  6. [14]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)

    Jiang, L., Mao, Y., Wang, X., Chen, X., Li, C.: Cola-diff: Conditional latent diffu- sion model for multi-modal mri synthesis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 398–408 (2023)

    2023

  7. [15]

    IEEE journal of biomedical and health informatics23(2), 538–546 (2018) 10 J

    Kawahara, J., Daneshvar, S., Argenziano, G., Hamarneh, G.: Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE journal of biomedical and health informatics23(2), 538–546 (2018) 10 J. Fayyad et al

    2018

  8. [16]

    Biomedical Signal Processing and Control106, 107735 (2025)

    Kui, X., Liu, B., Sun, Z., Li, Q., Zhang, M., Liang, W., Zou, B.: Med-lvdm: Medical latent variational diffusion model for medical image translation. Biomedical Signal Processing and Control106, 107735 (2025)

    2025

  9. [17]

    Applied Soft Computing161, 111714 (2024)

    Prusty, M.R., Sudharsan, R.M., Anand, P.: Enhancing medical image classification with generative ai using latent denoising diffusion probabilistic model and wiener filtering approach. Applied Soft Computing161, 111714 (2024)

    2024

  10. [18]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)

    2022

  11. [19]

    arXiv preprint arXiv:2308.12453 (2023)

    Sagers, L.W., Diao, J.A., Melas-Kyriazi, L., Groh, M., Rajpurkar, P., Adam- son, A.S., Rotemberg, V., Daneshjou, R., Manrai, A.K.: Augmenting medical im- age classifiers with synthetic data from latent diffusion models. arXiv preprint arXiv:2308.12453 (2023)

    2023 arXiv

  12. [20]

    Advances in neural information processing systems35, 25278–25294 (2022)

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large- scale dataset for training next generation image-text models. Advances in neural information processing systems35, 2...

    2022

  13. [21]

    Scientific data 5(1), 1–9 (2018)

    Tschandl, P., Rosendahl, C., Kittler, H.: The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data 5(1), 1–9 (2018)

    2018

  14. [22]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classi- fication and localization of common thorax diseases. In: Proceedings of the IEEE conference on computer vision and patte...

    2017

  15. [23]

    In: Medical Imaging with Deep Learning (MIDL)

    Wu, J., Fu, R., Fang, H., Zhang, Y., Yang, Y., Xiong, H., Liu, H., Xu, Y.: Med- segdiff: Medical image segmentation with diffusion probabilistic model. In: Medical Imaging with Deep Learning (MIDL). pp. 1623–1639 (2024)

    2024

  16. [24]

    In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024

    Xie, J., Chen, X., Zhao, Y., Meng, Y., Zhao, H., Nguyen, A., Li, X., Zheng, Y.: Multi-disease detection in retinal images guided by disease causal estimation. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. vol. LNCS 15001, pp. 743–753. Springer N...

    2024

  17. [25]

    In: Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Yellapragada, S., Graikos, A., Prasanna, P., Kurc, T., Saltz, J., Samaras, D.: Pathldm: Text conditioned latent diffusion model for histopathology. In: Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5182–5191 (2024)

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.