Breast Cancer Detection in Thermographic Images via Diffusion-Based Augmentation and Nonlinear Feature Fusion

Jingyu Liu; M. Moein Esfahani; Sepehr Salem; Vince Calhoun

REVIEW 4 major objections 5 minor 23 references

Breast Cancer Detection in Thermographic Images via Diffusion-Based Augmentation and Nonlinear Feature Fusion

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that diffusion-model augmentation plus a fusion of deep and nonlinear boundary features makes breast-cancer classification in thermograms reach 98.0% accuracy and 98.1% sensitivity, and that both components contribute signi

desk verdict Interesting framework, but the evaluation protocol is under-specified enough that the 98% accuracy may reflect leakage rather than generalization. read the letter →

arxiv 2509.07277 v1 pith:4YUOBVQD submitted 2025-09-08 cs.CV cs.AI

Sepehr Salem , M. Moein Esfahani , Jingyu Liu , Vince Calhoun This is my paper · ORCID

classification cs.CVcs.AI

keywords breastcancerthermographydiffusionprobabilisticmodeldataaugmentationnonlinearfeaturesfeaturefusionXGBoostmedicalimagesynthesis

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 is trying to establish that data scarcity in breast thermography can be overcome by a class-conditioned diffusion model that generates synthetic tumor-region patches, and that a classifier trained on those patches with a hybrid feature set reaches 98.0% accuracy and 98.1% sensitivity. It argues that diffusion-based augmentation beats affine transformations and a ProGAN baseline, and that adding four nonlinear boundary-complexity descriptors to ResNet-50 deep features produces a statistically significant gain. If the result holds, non-invasive thermal imaging becomes a more credible basis for accurate computer-aided breast cancer screening without requiring large annotated datasets.

What carries the argument

The load-bearing machinery is a two-branch feature pipeline built around a U-Net that segments the tumor ROI from each thermogram. The ROI feeds a class-conditioned Diffusion Probabilistic Model—a denoising generative network trained to reverse Gaussian corruption—that creates synthetic patches, and also feeds a frozen 2048-dimensional ResNet-50 feature extractor. In parallel, the segmented tumor boundary is turned into a radial-distance signal and summarized by four nonlinear descriptors: Lyapunov exponent, largest Lyapunov exponent, approximate entropy, and box-counting dimension. Concatenating the 2048 deep values with the 4 nonlinear values yields a 2052-dimensional fused vector that an

What would settle it

Fix the five folds before training the diffusion model, generate synthetic patches only from training-fold real ROIs, and rerun the exact pipeline; if accuracy falls from 98.0% toward the 95.5% deep-features-only level, leakage explains the headline. A complementary test is to retrain on DMR-IR and evaluate on an external multi-center thermogram set, where sensitivity markedly below 98.1% would show overfitting.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a class-conditioned diffusion model can synthesize realistic tumor-region thermogram patches—1,000 evenly split between benign and malignant—and that these synthetic patches, added to the original real images, let a frozen ResNet-50 plus four handcrafted descriptors reach 98.0% accuracy, 98.1% sensitivity, and 97.9% specificity in 5-fold cross-validation. The ablation path is explicit: real images with deep features alone give 90.1%; affine augmentation gives 92.5%; ProGAN augmentation gives 93.4%; DPM augmentation gives 95.5%; fusing the deep features with Lyapunov exponent, largest Lyapunov exponent, approximate entropy, and box-count

Load-bearing premise

The 98.0% result assumes the diffusion model is trained on training-fold real images only and that no synthetic image derived from a validation-fold real thermogram enters the training set.

Editorial extensions

If this is right

Diffusion-based augmentation displaces GAN-based augmentation in this setting: on the same ROI data, the DPM reaches FID 14.3 versus 25.8 and sFID 7.8 versus 15.2, and the accuracy gain over ProGAN is 95.5% versus 93.4%.
The four nonlinear descriptors carry information the deep features miss; appending 4 numbers to the 2048-dim vector moves accuracy from 95.5% to 98.0%.
Because the DPM is class-conditioned, the same pipeline can oversample whichever class is scarce, which is a direct remedy for class imbalance as well as overall data scarcity.
A frozen, pretrained feature extractor suffices, so the classification stage is cheap and does not require fine-tuning on the small medical dataset.

Reading between the lines

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

Implicit in the paper is a strict ordering requirement that the paper never states: the diffusion model must be trained on training-fold real ROIs before any synthetic image is added, or validation-fold information leaks into training. A leakage-safe rerun is the first thing to test.
If that ordering holds, the recipe should transfer to other small medical imaging tasks where a segmentation mask is available, because diffusion training and four scalar descriptors do not depend on thermography-specific assumptions.
One assumption worth probing is that the U-Net mask is accurate on synthetic images; since the nonlinear features are computed from those masks, segmentation error on generated images would dilute and possibly explain part of the nonlinear-feature gain.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's my take. The paper's actual novelty is narrow but real: it's the first to apply diffusion-based augmentation to breast thermograms, and it couples that with handcrafted nonlinear features (LE, ApEn, FD) that have been used elsewhere in medical imaging. The framework is described clearly, the ablations are structured, and the DPM's generative quality beats the ProGAN baseline on standard metrics. That part is fine.

The problem is the evaluation. The authors never tell us how big the dataset is, how the 5-fold cross-validation was constructed relative to the augmentation, or how the statistical tests were run. The phrase 'evaluated using cross-validation on the augmented dataset' (Section II-E) is exactly the kind of ambiguity that can hide leakage. If the DPM was trained on all real ROIs and synthetic patches were added before splitting, then the generative model has already seen the validation images, and synthetic patches derived from that distribution can end up in training folds. The same applies to the U-Net used for segmentation and nonlinear features. That could plausibly account for the 2.5-point jump (95.5 to 98.0) when nonlinear features are added. It may not be happening, but the paper gives no evidence that it isn't. The missing dataset size and statistical test details reinforce the concern.

There are also smaller issues: the ResNet-50 is frozen, which is sensible, but the choice of XGBoost over a simple classifier is not justified; the comparison set is thin (only affine, ProGAN, and no external state-of-the-art); and no code or data is released.

If the evaluation protocol is as clean as the abstract implies, the 98% result would be a useful data point. As written, the paper is not reproducible and the central number is not trustworthy without clarification. I would send it to peer review, because the idea deserves a careful referee who can force the authors to specify per-fold augmentation and ideally rerun the experiments with proper isolation. But I'm not citing it yet, and I'd use it in a reading group specifically to talk about evaluation leakage.

Referee Report

4 major / 5 minor

Summary. The paper proposes a breast cancer classification pipeline for thermographic images. A U-Net segments tumor ROIs; a diffusion probabilistic model (DPM) generates 1000 class-conditioned synthetic ROI patches; ResNet-50 deep features and four handcrafted nonlinear features (LE, LLE, ApEn, BCD/FD) are concatenated and fed to an XGBoost classifier. On the DMR-IR dataset, the authors report 98.0% accuracy, 98.1% sensitivity, and 97.9% specificity under 5-fold cross-validation, with ablations showing DPM augmentation and nonlinear feature fusion each improve accuracy (95.5% vs. 90.1% baseline; 98.0% after fusion). They also report FID 14.3 for the DPM versus 25.8 for ProGAN. The central claim is that the DPM augmentation and handcrafted feature fusion are responsible for the high accuracy.

Significance. If the reported results are sound, the paper would make a useful contribution to breast thermography CAD: it provides a concrete DPM-based augmentation pipeline, a comparison against a GAN baseline, and a hybrid deep/nonlinear feature representation. The idea of using diffusion models for thermogram augmentation is timely, and the inclusion of interpretable nonlinear features is a strength. However, the empirical evaluation as written does not establish the headline claim. The paper omits the dataset size and the exact cross-validation protocol and, more seriously, it does not state whether the DPM and U-Net are trained inside or outside the cross-validation folds. Because Table I's accuracy improvements (95.5% to 98.0%) are modest and could plausibly arise from information leakage, the central result is not yet reliable. The manuscript also reports statistical significance (p<0.01) without naming the test or providing the necessary details. These issues are fixable in principle but require either clarification or re-analysis with a leakage-free protocol.

major comments (4)

[Section IV, Section III-D] The evaluation protocol is not defined precisely enough to rule out information leakage. Equation (8) defines Xaug as the union of real and synthetic ROIs, and Section III-D says results are based on 5-fold cross-validation 'on the augmented dataset.' The paper never states whether the DPM and U-Net are trained before or after the train/validation split. The statement in Section III-E that 'both models were trained on the same dataset of real ROI patches' suggests the DPM is trained on the full real dataset. If so, for a validation real image x_v, a synthetic image generated from x_v can appear in the training fold; the classifier sees near-duplicates of validation samples during training. The same mechanism affects the nonlinear features because the U-Net was trained on real images and is applied to synthetic images derived from validation real images. This could explain the 2.5-point j
[Section IV, Section III-D] The claim that the improvements are 'statistically significant (p<0.01)' is unsupported. The manuscript does not name the statistical test (e.g., paired t-test, McNemar, Wilcoxon), does not state what is being compared (per-fold accuracy vectors? bootstrap resamples?), and does not provide the test statistic or the number of samples. Given that Table I reports means ± standard deviations over 5 folds, with only 5 folds a p<0.01 claim requires a very strong and well-specified test. This is load-bearing because the abstract and conclusion both rely on significance to justify the 'critical' role of the two proposed components.
[Section III-A, Table I] The dataset size and class composition are omitted. The paper says a 'balanced subset of benign and malignant cases' from DMR-IR was selected, but does not state the number of patients, images, or ROI patches, nor the number per class. The DPM is reported to generate 1000 synthetic ROIs evenly split, but without knowing the real dataset size the augmentation ratio is unknown. This prevents interpretation of Table I and of the claim that data scarcity is being addressed. The authors must disclose N, class balance, and whether cross-validation is at the image or patient level.
[Section II-D, Section III-E] The nonlinear features extracted from DPM-generated images are assumed meaningful, but this is not validated. A U-Net trained on real thermograms is applied to synthetic ROIs to obtain tumor masks, and the handcrafted features are computed from those masks. If the DPM generates images that are out of distribution, or if the DPM merely memorizes real ROIs, the segmentations may be unreliable. The manuscript does not report any quantitative evaluation of segmentation quality on synthetic images (e.g., Dice against manual masks, or a sanity check such as comparing feature distributions on real vs. synthetic images). Since the final accuracy gain in Table I comes exactly from these features, the validity of the nonlinear features on synthetic data is load-bearing.

minor comments (5)

[Section II-D, Eq. (12)] The acronym 'BCD' is introduced but Eq. (12) defines a fractal dimension 'D' and the text later refers to 'FD'. Use one consistent notation throughout; BCD and FD appear to be the same quantity.
[Section II-B, Eq. (7)] Equation (7) writes x0 on the left-hand side, but the denoising step in a DDPM should produce x_{t-1}, not x0. The notation is confusing and should be corrected or clarified.
[Section III-B, Table II] FID and sFID are computed on some set of real and synthetic images, but the manuscript does not state how many real and synthetic images were used for the metrics, whether the same set was used for ProGAN and DPM, or whether any repeated runs were performed. Add this information.
[Section III-A] The preprocessing steps (resize, normalization) are stated, but not the temperature scaling or any cropping details. For reproducibility, describe how ROIs are extracted from the U-Net mask (bounding box margins, minimum size, etc.).
[Section II-C] The paper says ResNet-50 is 'frozen', but Section II-C also says 'without risking overfitting on our smaller medical dataset.' Clarify whether the 2048-d features are extracted once and then used as inputs, or whether the model is trained in any way.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported accuracies are empirical results, not derived from self-citations or fitted definitions.

full rationale

The paper's central claims are empirical classification results: XGBoost trained on fused ResNet-50 and nonlinear features achieves 98.0% accuracy in 5-fold cross-validation. These numbers are experimental outcomes, not quantities derived from a fitted equation or from prior work by the authors. The augmentation set in Eq. (8) is an input to the classifier, not a construction that forces the reported accuracy. Self-citations (e.g., [5], [14], [15]) are used for motivation and context, but the paper's comparative evaluation of DPM versus ProGAN and the ablation of nonlinear features is self-contained experimental evidence. The potential data-leakage issue (DPM or U-Net trained before splitting, synthetic copies of validation real images entering training) is a real methodological validity concern, but it is not circularity in the sense of a prediction reducing to its inputs by construction; it is a possible contamination of the evaluation protocol. The paper also acknowledges the need for external multicenter validation, further indicating that the reported results are empirical and not claimed to be first-principles derivations. Therefore, no circular step is exhibited under the required standard.

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

No new physical or mathematical entities are introduced. The central results rest on hand-chosen parameters (synthetic image count, ApEn/LE settings, XGBoost settings) and on domain assumptions about the validity of nonlinear features and the transferability of U-Net segmentation to synthetic images.

free parameters (6)

Number of synthetic ROI patches = 1000
Hand-chosen augmentation budget, evenly split between benign and malignant; directly affects training set composition and reported accuracy.
ApEn embedding dimension m = 2
Set to 2 with tolerance r=0.2*std; standard values but affect feature values.
ApEn tolerance r = 0.2*SD
Chosen from literature; affects regularity measure.
LE embedding dimension and time delay = dim=3, delay=1
Hand-chosen for phase space reconstruction of radial contour signal.
XGBoost hyperparameters = not reported
No details on learning rate, max depth, n_estimators, etc.; likely tuned but undisclosed.
DPM training configuration = 1000 steps, linear schedule, 200k iters, batch 32, AdamW LR 1e-4
Standard choices, but could affect synthetic image quality and downstream accuracy.

assumptions (5)

standard math The diffusion forward and reverse processes as defined by Ho et al. (2020) generate valid samples.
The paper builds on standard DDPM theory; no derivation is given, so it relies on published results.
domain assumption Nonlinear features (LE, ApEn, BCD) computed from tumor contour radial signals capture malignancy-related complexity.
The paper asserts this connection based on prior literature (refs 12-18) but provides no independent validation in this dataset.
domain assumption A U-Net trained on real thermograms produces accurate tumor masks for DPM-generated synthetic images.
The nonlinear features for synthetic images depend on this transfer; no evaluation of segmentation quality on synthetic images is provided.
domain assumption DMR-IR dataset labels are correct and the chosen balanced subset is representative.
The paper provides no details on the selection or label verification.
ad hoc to paper Synthetic images are equivalent to real images for training the classifier.
The framework treats DPM-generated ROIs as real training data; this is the core augmentation assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Breast Cancer Detection in Thermographic Images via Diffusion-Based Augmentation and Nonlinear Feature Fusion." pith.science (2026). https://pith.science/paper/4YUOBVQD

@misc{pith2026250907277,
  author       = {Pith},
  title        = {Pith review of: Breast Cancer Detection in Thermographic Images via Diffusion-Based Augmentation and Nonlinear Feature Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4YUOBVQD}},
  note         = {Machine review of arXiv:2509.07277}
}

read the original abstract

Data scarcity hinders deep learning for medical imaging. We propose a framework for breast cancer classification in thermograms that addresses this using a Diffusion Probabilistic Model (DPM) for data augmentation. Our DPM-based augmentation is shown to be superior to both traditional methods and a ProGAN baseline. The framework fuses deep features from a pre-trained ResNet-50 with handcrafted nonlinear features (e.g., Fractal Dimension) derived from U-Net segmented tumors. An XGBoost classifier trained on these fused features achieves 98.0\% accuracy and 98.1\% sensitivity. Ablation studies and statistical tests confirm that both the DPM augmentation and the nonlinear feature fusion are critical, statistically significant components of this success. This work validates the synergy between advanced generative models and interpretable features for creating highly accurate medical diagnostic tools.

Figures

Figures reproduced from arXiv: 2509.07277 by the authors.

**Figure 1.** Overview of the proposed hybrid framework for breast cancer classification. First, a U-Net model segments the tumor [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Tumor boundary converted to radial distance time series. Regular contours (a) produce smooth signals; irregular contours [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Qualitative comparison of synthetic thermogram ROI [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages

[1]

Khodadadi and S

H. Khodadadi and S. Nazem, ”Improving cancer detection through computer-aided diagnosis: A comprehensive analysis of nonlinear and texture features in breast thermograms,”PLoS One, vol. 20, no. 5, 2025, p. e0322934

work page 2025
[2]

E. A. Mohamedet al., ”Deep learning model for fully automated breast cancer detection system from thermograms,”PLoS One, vol. 17, no. 1, 2022, p. e0262349

work page 2022
[3]

M. A. S. Al Husaini, M. H. Habaebi, and M. R. Islam, ”Real-time thermography for breast cancer detection with deep learning,”Discover Artif. Intell., vol. 4, no. 1, 2024, p. 57

work page 2024
[4]

M ¨uller-Franzeset al., ”A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis,”Sci

G. M ¨uller-Franzeset al., ”A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis,”Sci. Rep., vol. 13, no. 1, 2023, p. 12098

work page 2023
[5]

Deep learning for automated detection of breast cancer in deep ultraviolet fluorescence images with diffusion probabilistic model,

S. S. Ghahfarrokhiet al., “Deep learning for automated detection of breast cancer in deep ultraviolet fluorescence images with diffusion probabilistic model,” inProc. IEEE Int. Symp. Biomed. Imaging (ISBI), 2024

work page 2024
[6]

J. Ho, A. Jain, and P. Abbeel, ”Denoising diffusion probabilistic models,” inAdvances in Neural Inf. Process. Syst., vol. 33, 2020, pp. 6840–6851

work page 2020
[7]

Choiet al., ”Perception prioritized training of diffusion models,” in Proc

J. Choiet al., ”Perception prioritized training of diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022

work page 2022
[8]

Veerlapalli and S

P. Veerlapalli and S. R. Dutta, ”A hybrid GAN-based deep learning framework for thermogram-based breast cancer detection,”Sci. Rep., vol. 15, no. 1, 2025, pp. 1–33

work page 2025

Show all 23 references

[9]

P. A. Moghadamet al., ”A morphology focused diffusion probabilistic model for synthesis of histopathology images,” inProc. IEEE/CVF Winter Conf. Appl. Comput. Vis., 2023

2023
[10]

Zhou and H

Q. Zhou and H. Yin, ”A U-Net based progressive GAN for microscopic image augmentation,” inProc. Med. Image Understand. Anal., Springer, 2022, pp. 458–468

2022
[11]

Kim and H

J. Kim and H. Park, ”Adaptive latent diffusion model for 3D medical image to image translation: Multi-modal magnetic resonance imaging study,”arXiv preprint, arXiv:2311.00265, 2023

2023 arXiv
[12]

N. S. Aghdamet al., ”Designing and comparing different color map algorithms for pseudo-coloring breast thermograms,”J. Med. Imaging Health Inform., vol. 3, no. 4, 2013, pp. 487–493

2013
[13]

Mazaheri and H

V . Mazaheri and H. Khodadadi, ”Heart arrhythmia diagnosis based on the combination of morphological, frequency and nonlinear features of ECG signals and metaheuristic feature selection algorithm,”Expert Syst. Appl., vol. 161, 2020, p. 113697

2020
[14]

S. S. Ghahfarrokhiet al., ”Malignant melanoma diagnosis applying a machine learning method based on the combination of nonlinear and texture features,”Biomed. Signal Process. Control, vol. 80, 2023, p. 104300

2023
[15]

S. S. Ghahfarrokhi and H. Khodadadi, ”Human brain tumor diagnosis using the combination of the complexity measures and texture features through magnetic resonance image,”Biomed. Signal Process. Control, vol. 61, 2020, p. 102025

2020
[16]

Khodadadiet al., ”Nonlinear analysis of the contour boundary irregularity of skin lesion using Lyapunov exponent and KS entropy,”J

H. Khodadadiet al., ”Nonlinear analysis of the contour boundary irregularity of skin lesion using Lyapunov exponent and KS entropy,”J. Med. Biol. Eng., vol. 37, 2017, pp. 409–419

2017
[17]

Khodadadiet al., ”Applying a modified version of Lyapunov ex- ponent for cancer diagnosis in biomedical images: the case of breast mammograms,”Multidimens

H. Khodadadiet al., ”Applying a modified version of Lyapunov ex- ponent for cancer diagnosis in biomedical images: the case of breast mammograms,”Multidimens. Syst. Signal Process., vol. 29, no. 1, 2018, pp. 19–33

2018
[18]

Arab Zade and H

M. Arab Zade and H. Khodadadi, ”Fuzzy controller design for breast cancer treatment based on fractal dimension using breast thermograms,” IET Syst. Biol., vol. 13, no. 1, 2019, pp. 1–7

2019
[19]

N. V . Shree and T. N. R. Kumar, ”Identification and classification of brain tumor MRI images with feature extraction using DWT and probabilistic neural network,”Brain Inform., vol. 5, no. 1, 2018, pp. 23–30

2018
[20]

A new database for breast research with infrared image,

L. Silvaet al., “A new database for breast research with infrared image,” J. Med. Imaging Health Inf., vol. 4, no. 1, pp. 92–100, 2014

2014
[21]

Improved techniques for training GANs,

T. Salimanset al., “Improved techniques for training GANs,” inAd- vances in Neural Inf. Process. Syst., vol. 29, 2016

2016
[22]

The role of ImageNet classes in Fr ´echet Inception Distance,

T. Kynk ¨a¨anniemiet al., “The role of ImageNet classes in Fr ´echet Inception Distance,”arXiv preprint, arXiv:2203.06026, 2022

2022 arXiv
[23]

Rethinking the Inception architecture for computer vision,

C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception architecture for computer vision,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 2818–2826

2016

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Khodadadi and S

H. Khodadadi and S. Nazem, ”Improving cancer detection through computer-aided diagnosis: A comprehensive analysis of nonlinear and texture features in breast thermograms,”PLoS One, vol. 20, no. 5, 2025, p. e0322934

work page 2025

[2] [2]

E. A. Mohamedet al., ”Deep learning model for fully automated breast cancer detection system from thermograms,”PLoS One, vol. 17, no. 1, 2022, p. e0262349

work page 2022

[3] [3]

M. A. S. Al Husaini, M. H. Habaebi, and M. R. Islam, ”Real-time thermography for breast cancer detection with deep learning,”Discover Artif. Intell., vol. 4, no. 1, 2024, p. 57

work page 2024

[4] [4]

M ¨uller-Franzeset al., ”A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis,”Sci

G. M ¨uller-Franzeset al., ”A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis,”Sci. Rep., vol. 13, no. 1, 2023, p. 12098

work page 2023

[5] [5]

Deep learning for automated detection of breast cancer in deep ultraviolet fluorescence images with diffusion probabilistic model,

S. S. Ghahfarrokhiet al., “Deep learning for automated detection of breast cancer in deep ultraviolet fluorescence images with diffusion probabilistic model,” inProc. IEEE Int. Symp. Biomed. Imaging (ISBI), 2024

work page 2024

[6] [6]

J. Ho, A. Jain, and P. Abbeel, ”Denoising diffusion probabilistic models,” inAdvances in Neural Inf. Process. Syst., vol. 33, 2020, pp. 6840–6851

work page 2020

[7] [7]

Choiet al., ”Perception prioritized training of diffusion models,” in Proc

J. Choiet al., ”Perception prioritized training of diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022

work page 2022

[8] [8]

Veerlapalli and S

P. Veerlapalli and S. R. Dutta, ”A hybrid GAN-based deep learning framework for thermogram-based breast cancer detection,”Sci. Rep., vol. 15, no. 1, 2025, pp. 1–33

work page 2025