arxiv: 2605.11758 · v1 · submitted 2026-05-12 · 📡 eess.IV · cs.CV

Recognition: 2 theorem links

· Lean Theorem

DiffSegLung: Diffusion Radiomic Distillation for Unsupervised Lung Pathology Segmentation

Catalin Fetita, Pierre-Yves Brillet, Rezkellah Noureddine Khiati

Pith reviewed 2026-05-13 05:15 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords unsupervised segmentationdiffusion modelslung CTradiomicspathology segmentationcontrastive learningHounsfield units

0 comments

The pith

Handcrafted radiomic descriptors can guide a diffusion model to segment lung pathologies in CT scans without annotations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to perform unsupervised segmentation of multiple lung pathologies in CT images by distilling information from radiomic descriptors into a diffusion network. These descriptors capture tissue differences based on density values and act as a teacher signal during training via a contrastive objective. At test time the network's internal features are clustered to produce labels, with boundary refinement added. A sympathetic reader would care because the lack of multi-pathology annotations currently blocks scalable analysis of routine scans. If the approach holds, large unlabeled CT collections could yield pathology maps directly from the model's learned representations.

Core claim

Diffusion Radiomic Distillation uses handcrafted radiomic descriptors as a physics-grounded teacher to shape the bottleneck of a 3D diffusion U-Net via a contrastive objective, transferring pathology-discriminative structure into the learned representation without annotations. At inference the teacher is discarded and multitimestep bottleneck features are clustered by a Gaussian Mixture Model with HU-guided label assignment, followed by boundary refinement. On 190 expert-annotated axial slices from four heterogeneous CT cohorts this yields improved segmentation across four pathology classes over unsupervised baselines and higher generation fidelity than prior CT diffusion models.

What carries the argument

Diffusion Radiomic Distillation: radiomic descriptors serve as a teacher to shape the diffusion U-Net bottleneck through a contrastive objective, enabling later clustering of the resulting features.

If this is right

Segmentation performance improves for all four tested pathology classes on data drawn from multiple CT sources.
The diffusion model itself generates higher-fidelity CT images than previous unsupervised approaches.
The radiomic teacher can be removed at test time, leaving only feature clustering and refinement.
The same trained features support consistent labeling across heterogeneous acquisition protocols without retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The distillation step could be tried on other density-sensitive imaging tasks such as liver or bone lesion segmentation where similar physical descriptors exist.
If the clustered features prove stable, they might later support downstream clinical predictions like progression risk without new labels.
Extending the contrastive objective to include additional physical signals beyond density could further reduce reliance on handcrafted descriptors.

Load-bearing premise

Handcrafted radiomic descriptors based on density values reliably encode the differences between pathology classes so that contrastive training plus clustering can separate them correctly without any labels or manual tuning.

What would settle it

Run the trained model on a fresh collection of expert-annotated CT slices from a new scanner or patient group and measure whether the segmentation overlap scores for each pathology class fail to exceed those obtained by clustering raw intensity values or other simple unsupervised baselines.

Figures

Figures reproduced from arXiv: 2605.11758 by Catalin Fetita, Pierre-Yves Brillet, Rezkellah Noureddine Khiati.

**Figure 1.** Figure 1: Training: ϵθ is trained on HU-preserved patches via Ldiff ; radiomic descriptors r ∈ R 34 align student embeddings via InfoNCE (discarded at inference). Inference: multi-timestep features fseg are clustered by GMM and refined by Sobel-Diffusion. Diffusion student. The bottleneck f (i) is projected via ϕs : R 256 → S128: z (i) diff = ℓ2 V2 σ V1 Flatten AvgPool3D(f (i) ) . (2) Radiomic similarity as… view at source ↗

**Figure 2.** Figure 2: Qualitative generation comparison (random axial slices). From top: real HUpreserved CT, LungDDPM+ [7], MedVAE [12], and DiffSegLung (ours). Our model preserves HU density ranges and parenchymal texture more faithfully than both baselines [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative segmentation comparison on representative axial slices from the 190-slice evaluation set [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

read the original abstract

Unsupervised segmentation of pulmonary pathologies in CT remains an open challenge due to the absence of annotated multi pathology cohorts and the failure of existing diffusion-based methods to exploit the quantitative Hounsfield Unit (HU) signal that physically distinguishes tissue classes. To address this, we propose DiffSegLung,a framework that introduces Diffusion Radiomic Distillation, in which handcrafted radiomic descriptors serve as a physics grounded teacher to shape the bottleneck of a 3D diffusion U-Net via a contrastive objective, transferring pathology discriminative structure into the learned representation without any annotations. At inference, the teacher is discarded and multitimestep bottleneck features are clustered by a Gaussian Mixture Model with HU-guided label assignment, followed by Sobel Diffusion Fusion for boundary refinement. Evaluated on 190 expert annotated axial slices drawn from four heterogeneous CT cohorts, Diff-SegLung improves segmentation across all four pathology classes over unsupervised baselines and improves generation fidelity over prior CT diffusion models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DiffSegLung distills handcrafted radiomics into a diffusion U-Net bottleneck via contrastive loss for unsupervised CT pathology segmentation, but the abstract supplies no metrics or ablations so the gains cannot be checked.

read the letter

The paper's main contribution is a training-only teacher that uses radiomic descriptors to shape the bottleneck features of a 3D diffusion U-Net through contrastive learning, after which the teacher is dropped and multi-timestep features are clustered by GMM with HU-guided assignment plus Sobel refinement. This produces unsupervised segmentations for four lung pathology classes on CT. The approach is new in its specific combination of radiomic distillation, diffusion bottleneck, and HU-guided clustering, and it tries to exploit the quantitative nature of Hounsfield units rather than treating the images as generic intensity data. Evaluating on 190 slices from four heterogeneous cohorts is a reasonable step toward showing some robustness beyond single-center data, and the claim of improved generation fidelity over prior CT diffusion models is at least a concrete comparison point. The framework description is technically coherent and keeps inference free of the teacher, which is a practical plus. The central assumption that the chosen radiomics reliably separate the pathology classes across varied scanners is the weakest link, and the abstract gives no evidence that the contrastive step actually transfers that structure or that the GMM step avoids post-hoc choices. No quantitative Dice or IoU numbers, no error bars, no ablation results, and no details on data splits or hyperparameter selection appear in the provided text, so the performance claims stay unverified. The stress-test concern about radiomics failing to provide sufficient discriminative signal holds up on the abstract alone. This work is aimed at researchers building unsupervised methods for medical CT analysis, especially those already using diffusion models. Readers looking for concrete ways to inject domain-specific radiomic or HU information into generative representations might pick up usable ideas. It deserves peer review because the problem is real and the pipeline is clearly laid out, but the authors will need to add the missing experiments and direct tests of the radiomics assumption before the claims can be taken seriously.

Referee Report

3 major / 2 minor

Summary. The paper introduces DiffSegLung, a framework for unsupervised segmentation of lung pathologies in CT images. It employs Diffusion Radiomic Distillation where handcrafted radiomic descriptors serve as a physics-grounded teacher to guide the bottleneck features of a 3D diffusion U-Net using a contrastive objective. At inference, multi-timestep features are clustered using a Gaussian Mixture Model with HU-guided label assignment and refined with Sobel Diffusion Fusion. The method is claimed to outperform unsupervised baselines on 190 expert-annotated axial slices from four heterogeneous CT cohorts across four pathology classes and to improve generation fidelity over prior CT diffusion models.

Significance. If the results hold, this work could be significant for the field of medical image segmentation by providing an unsupervised method that leverages quantitative HU information and radiomics to handle multiple pathologies without annotations. The distillation approach from radiomics to diffusion models is innovative and addresses a key challenge in pulmonary CT analysis where labeled data is limited. The evaluation on heterogeneous cohorts adds to its potential applicability.

major comments (3)

[Abstract] Abstract: The abstract claims performance improvements over unsupervised baselines but provides no quantitative metrics, error bars, ablation studies, or details on data splits and hyperparameter selection. This undermines the ability to assess the central claim of superiority, as the evaluation is load-bearing for the paper's contribution.
[Method] Method (Diffusion Radiomic Distillation): The core assumption that handcrafted radiomic descriptors reliably encode pathology-discriminative structure based on HU values is not sufficiently validated. No evidence or analysis is presented showing that these features separate the four pathology classes across the heterogeneous cohorts, which is critical for the contrastive objective to transfer useful structure and for the subsequent GMM clustering to succeed without supervision.
[Experiments] Experiments: The evaluation on 190 slices from four cohorts lacks details on how the cohorts were split for training and testing, the specific pathology classes, and statistical comparisons. Without these, the claim of improvement across all classes cannot be properly evaluated.

minor comments (2)

[Abstract] There is an inconsistency in the naming: 'DiffSegLung' and 'Diff-SegLung' are used interchangeably.
[Abstract] The term 'multitimestep' should be hyphenated as 'multi-timestep' for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We have carefully reviewed each major comment and provide point-by-point responses below. We agree that several clarifications and additions will strengthen the paper and will incorporate them in the revised version.

read point-by-point responses

Referee: [Abstract] Abstract: The abstract claims performance improvements over unsupervised baselines but provides no quantitative metrics, error bars, ablation studies, or details on data splits and hyperparameter selection. This undermines the ability to assess the central claim of superiority, as the evaluation is load-bearing for the paper's contribution.

Authors: We agree that the abstract would benefit from including key quantitative results to better support the central claims. In the revised manuscript, we will update the abstract to report mean Dice scores (with standard deviations) for each pathology class, along with a brief reference to the cross-cohort evaluation protocol. Ablation studies and hyperparameter details remain in the main Experiments section and supplementary material, but we will add a concise mention of the evaluation setup in the abstract for improved readability. revision: yes
Referee: [Method] Method (Diffusion Radiomic Distillation): The core assumption that handcrafted radiomic descriptors reliably encode pathology-discriminative structure based on HU values is not sufficiently validated. No evidence or analysis is presented showing that these features separate the four pathology classes across the heterogeneous cohorts, which is critical for the contrastive objective to transfer useful structure and for the subsequent GMM clustering to succeed without supervision.

Authors: We acknowledge that an explicit validation of the radiomic features' ability to discriminate pathology classes would strengthen the methodological justification. We will add a new subsection or figure in the revised manuscript (e.g., t-SNE embeddings or class-separation metrics such as silhouette scores computed on the radiomic descriptors) demonstrating separation across the four classes and heterogeneous cohorts. This analysis will directly support the rationale for the contrastive distillation and the unsupervised GMM step. revision: yes
Referee: [Experiments] Experiments: The evaluation on 190 slices from four cohorts lacks details on how the cohorts were split for training and testing, the specific pathology classes, and statistical comparisons. Without these, the claim of improvement across all classes cannot be properly evaluated.

Authors: We agree that additional experimental details are required for full reproducibility and assessment. In the revised Experiments section, we will explicitly describe the training/testing splits (e.g., leave-one-cohort-out or stratified per-cohort ratios), list the four pathology classes, and report statistical comparisons (paired t-tests or Wilcoxon signed-rank tests with p-values) between DiffSegLung and baselines. These additions will directly address the evaluation of improvements across classes. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external handcrafted radiomics and standard clustering

full rationale

The paper introduces handcrafted radiomic descriptors (based on HU values) as an external physics-grounded teacher for contrastive distillation into the 3D diffusion U-Net bottleneck. At inference the teacher is dropped and multitimestep features are clustered via GMM with HU-guided assignment. No equations or steps reduce the central claim to a quantity defined by parameters fitted inside the same loop, nor do they rely on self-citation chains or ansatzes smuggled from prior author work. The radiomics and GMM are presented as independent inputs, making the chain self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that handcrafted radiomics capture pathology-discriminative HU-based structure that can be transferred via contrastive learning; no free parameters or invented entities are explicitly quantified in the abstract.

axioms (1)

domain assumption Handcrafted radiomic descriptors provide physics-grounded, pathology-discriminative information based on Hounsfield Unit values.
Invoked when radiomics are positioned as the teacher that shapes the diffusion bottleneck.

pith-pipeline@v0.9.0 · 5473 in / 1264 out tokens · 59473 ms · 2026-05-13T05:15:28.977009+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
handcrafted radiomic descriptors serve as a physics grounded teacher... 34-dimensional radiomic vector r(i) comprising GLCM statistics (14 features), LBP histograms (8 features), Gabor filter responses (8 features), and first-order HU statistics (4 features)
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear
GMM with K=5 components... HU-guided label assignment... comparing each cluster’s mean HU h̄k against clinically established thresholds

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · 1 internal anchor

[1]

Radiology307(5), e222998 (2023)

Amudala Puchakayala, P.R., Sthanam, V.L., Nakhmani, A., et al.: Radiomics for improved detection of chronic obstructive pulmonary disease in low-dose and standard-dose chest CT scans. Radiology307(5), e222998 (2023)

work page 2023
[2]

Bandara, W.G.C., Nair, N.G., Patel, V.M.: DDPM-CD: Remote sensing change detection using denoising diffusion probabilistic models (2022)

work page 2022
[3]

Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label- efficient semantic segmentation with diffusion models (2022)

work page 2022
[4]

Radi- ology290(3), 783–792 (2019)

Beig, N., Khorrami, M., Alilou, M., et al.: Perinodular and intranodular radiomic features on lung CT images distinguish adenocarcinomas from granulomas. Radi- ology290(3), 783–792 (2019)

work page 2019
[5]

Denoising Diffusion Probabilistic Models

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. CoRR abs/2006.11239(2020),https://arxiv.org/abs/2006.11239

work page internal anchor Pith review Pith/arXiv arXiv 2006
[6]

IEEE Transactions on Biomedical Engineering73(3), 1134–1145 (2026)

Jiang, Y., Lemaréchal, Y., Plante, S., Bafaro, J., Abi-Rjeile, J., Joubert, P., De- sprés, P., Manem, V.: Lung-DDPM: Semantic layout-guided diffusion models for thoracic CT image synthesis. IEEE Transactions on Biomedical Engineering73(3), 1134–1145 (2026). https://doi.org/10.1109/tbme.2025.3599011

work page doi:10.1109/tbme.2025.3599011 2026
[7]

Computers in Biology and Medicine199, 111290 (2025)

Jiang, Y., Shariftabrizi, A., Manem, V.S.: Lung-DDPM+: Efficient thoracic CT image synthesis using diffusion probabilistic model. Computers in Biology and Medicine199, 111290 (2025). https://doi.org/10.1016/j.compbiomed.2025.111290

work page doi:10.1016/j.compbiomed.2025.111290 2025
[8]

Advances in Neural Information Processing Systems35, 5775–5787 (2022)

Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems35, 5775–5787 (2022)

work page 2022
[9]

Luo, X., Hu, M., Liao, W., Zhai, S., Song, T., Wang, G., Zhang, S.: Scribble- supervised medical image segmentation via dual-branch network and dynamically mixed pseudo labels supervision (2022)

work page 2022
[10]

IEEE Transactions on Medical Imaging40(12), 3820–3831 (2021)

Qin, D., Bu, J.J., Liu, Z., Shen, X., Zhou, S., Gu, J.J., Wang, Z.H., Wu, L., Dai, H.F.: Efficient medical image segmentation based on knowledge dis- tillation. IEEE Transactions on Medical Imaging40(12), 3820–3831 (2021). https://doi.org/10.1109/TMI.2021.3098703

work page doi:10.1109/tmi.2021.3098703 2021
[11]

Tian, J., Aggarwal, L., Colaco, A., Kira, Z., Gonzalez-Franco, M.: Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion (2024)

work page 2024
[12]

Varma, M., Kumar, A., van der Sluijs, R., et al.: MedVAE: Efficient automated interpretation of medical images with large-scale generalizable autoencoders (2025)

work page 2025
[13]

Wu, J., Fu, R., Fang, H., Zhang, Y., Yang, Y., Xiong, H., Liu, H., Xu, Y.: Med- SegDiff:Medicalimagesegmentationwithdiffusionprobabilisticmodel.In:Medical Imaging with Deep Learning. pp. 1623–1639. PMLR (2024)

work page 2024
[14]

Investigative Radiology 47, 596–602 (2012)

Zach, J.A., Newell, J.D., Schroeder, J.D., et al.: Quantitative computed tomogra- phy of the lungs and airways in healthy nonsmoking adults. Investigative Radiology 47, 596–602 (2012)

work page 2012
[15]

Computers in Biology and Medicine164, 107284 (2023)

Zhao, L., Qian, X., Guo, Y., Song, J., Hou, J., Gong, J.: MSKD: Structured knowledge distillation for efficient medical image segmen- tation. Computers in Biology and Medicine164, 107284 (2023). https://doi.org/10.1016/j.compbiomed.2023.107284

work page doi:10.1016/j.compbiomed.2023.107284 2023