ImmunoDiff: A Diffusion Model for Immunotherapy Response Prediction in Lung Cancer

Amna F. Sher; Chao Chen; Gagandeep Singh; Judy Huang; Moinak Bhattacharya; Prateek Prasanna

REVIEW 4 major objections 6 minor 48 references

ImmunoDiff: A Diffusion Model for Immunotherapy Response Prediction in Lung Cancer

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that immunotherapy response in non-small cell lung cancer is better predicted from a diffusion model's generated post-treatment CT than from the pre-treatment scan alone.

desk verdict Coherent, novel integration of anatomy-guided diffusion and clinical conditioning for post-treatment CT synthesis, but the evaluation protocol has a possible leakage flaw and the numbers are too fragile to trust at face value. read the letter →

arxiv 2505.23675 v1 pith:FCL5BQP5 submitted 2025-05-29 eess.IV cs.CV

Moinak Bhattacharya , Judy Huang , Amna F. Sher , Gagandeep Singh , Chao Chen , Prateek Prasanna This is my paper

classification eess.IVcs.CV

keywords diffusionmodelsimmunotherapyresponsepredictionnon-smallcelllungcancerCTimagesynthesisanatomy-guidedgenerationclinicalvariableconditioningsurvivalanalysisimage-to-imagetranslation

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

The paper tries to establish that immunotherapy response in non-small cell lung cancer is better predicted by a generative model that first synthesizes the post-treatment CT from the pre-treatment scan and clinical variables, and then reads response off the frozen diffusion model's features, than by any classifier trained directly on pre-treatment images. On a 74-patient cohort treated with immune checkpoint inhibitors, the approach reports 0.74±0.20 balanced accuracy, a 21.24% gain over the second-best baseline, and a c-index of 0.57±0.12 for survival. If this holds, physicians would gain a non-invasive, imaging-based forecast of whether a patient is likely to respond before measurable tumor shrinkage appears on a real follow-up scan.

What carries the argument

The load-bearing object is a two-stage diffusion pipeline with a frozen denoising U-Net at its center. Stage 1 pre-trains a standard denoising diffusion probabilistic model (DDPM) on public NSCLC CT data, then freezes it and trains a ControlNet—a conditioning side-branch with zero-initialized convolutions—that accepts embeddings from vessel and lobe masks as anatomical controls. Stage 2 adds the cbi-Adapter, a joint-attention module with rotary position embeddings that merges clinical variables (demographics, blood biomarkers, PD-L1 status) with the pre-treatment CT embedding and feeds the combined condition into the same frozen network to generate the post-treatment image. The prediction head is not a separate encoder: it consumes the up-block features of the frozen diffusion U-Net, so the same generative model that forecasts the follow-up scan also supplies the representation for response and survival.

What would settle it

A shuffle test—training the pipeline on pairs where post-treatment scans are randomly reassigned to different patients—would settle whether the model learns patient-specific treatment response, and a simpler check is the Dice overlap of lobe or vessel masks between pre- and post-treatment crops, since high misalignment would break the voxel-level translation premise.

Watch

Extended reading notes

Core claim

ImmunoDiff's central claim is that the morphological and textural transformation from pre- to post-treatment CT carries the answer to whether immunotherapy will work, and that diffusion models can both represent that transformation and serve as the feature extractor for prediction. The paper trains an anatomy-conditioned denoising diffusion model on large public lung-cancer CT collections, then fine-tunes it on paired pre/post scans from 74 patients with clinical variables as additional conditioning through a joint-attention adapter. From the same frozen U-Net it extracts up-block features for two downstream heads: responder-versus-non-responder classification and survival hazard prediction. The reported results—balanced accuracy 0.74±0.20, F1 0.76±0.13, c-index 0.57±0.12—beat direct classifiers, clinical-augmented classifiers, and diffusion-based baselines, and the ablations show that neither anatomy guidance alone nor clinical conditioning alone achieves the combined performance.

Load-bearing premise

The model assumes the pre-treatment and post-treatment CT scans are spatially aligned after cropping, so the diffusion model is learning a true anatomical transformation rather than compensating for misalignment.

Editorial extensions

If this is right

The generated post-treatment image becomes a candidate biomarker: it is a per-patient visual forecast of whether the tumor will shrink, stabilize, or grow under immunotherapy.
A frozen diffusion U-Net can serve as the representation extractor for both classification and survival, eliminating the need for a separately trained imaging encoder.
Pre-training on large public lung-cancer CT collections transfers to small immunotherapy cohorts, relieving the data bottleneck that limits direct supervised predictors.
Anatomical controls (lobes and vessels) contribute to generated-image fidelity, since ablating them degrades image quality and downstream prediction performance.
Clinical variables act as generative controls rather than mere classifier inputs: they shape the synthesized post-treatment anatomy.

Reading between the lines

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

A reader could test the same recipe in other solid tumors with paired pre/post treatment imaging, such as colorectal or melanoma, to see if the generative forecast transfers beyond NSCLC.
The cbi-Adapter is a generic multimodal fusion module; separating it from the diffusion setting and using it in a direct imaging-plus-clinical classifier would isolate how much of the gain comes from conditioning versus from the generative feature extractor.
Synthesized post-treatment images could be tested against longer-term endpoints (progression-free survival, durable benefit) to see whether the forecast captures durable response rather than just 6-week RECIST change.
A multi-center replication on a larger cohort is the natural next step; the reported margins are computed on 74 patients, so the generalizability of the 21% gain remains unmeasured.

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

Two things to know up front. First, the integration is genuinely new: nobody else has combined an anatomy-guided ControlNet with clinical variable conditioning to synthesize post-treatment CTs and then used the frozen U-Net features for immunotherapy response and survival prediction. Second, the evaluation is not yet solid enough to take the headline numbers seriously. The main question is whether the Stage 2 diffusion model is retrained inside each cross-validation fold or trained once on all 74 patients. If it is the latter, the test patient's post-treatment CT was already a generation target for the frozen feature extractor, and the features can leak the RECIST outcome directly. That would inflate the 0.74 balanced accuracy and 0.57 c-index independently of real generalization. The paper never says which it is, and that is a load-bearing ambiguity.

What the paper does well: the two-stage pretraining on public CT data is a sensible way to compensate for the tiny cohort. The cbi-Adapter is simple but does what it claims. The ablations are the right checks to run, and the writing is clear about the role of each stage. I believe the authors are not trying to deceive; the issue is sloppy protocol description, not bad faith.

Now the soft spots in proportion. The cohort is 74 patients with 19 responders, and the error bars are enormous: balanced accuracy 0.74±0.20. There are no significance tests, no bootstrap confidence intervals, and the c-index gain over the second-best baseline is 0.03. Table 1 has duplicate baseline rows (ViT-B/16 and Swin-Base share identical numbers, and several +CX rows match each other exactly), which looks like a copy-paste error and makes me question the care with which the results were assembled. The absence of any stated registration or alignment between pre- and post-treatment CTs is a separate concern; cropping to (96,96,16) without alignment makes the conditioning on z0 suspect, though it would not by itself inflate the prediction numbers as the leakage route would. No code or data has been released despite the promise, so nothing here is independently reproducible.

Net: this is a coherent, well-motivated integration paper with a fixable but load-bearing evaluation flaw. If leakage is ruled out, the result is a modest but real step for the subfield. If not, the headline numbers are meaningless. The paper deserves a serious referee, but it is not ready to be trusted at face value.

For whom? Researchers working on generative models for medical outcome prediction, not clinicians. My recommendation: send it to peer review with the expectation of heavy revision, and require an explicit statement of the cross-validation protocol plus code and fold splits before the 21% claim can be assessed.

Referee Report

4 major / 6 minor

Summary. The paper proposes ImmunoDiff, a two-stage diffusion framework for immunotherapy response prediction in NSCLC. Stage 1 pretrains a DDPM on public TCIA data with vessel/lobe anatomical conditioning via contrastively trained encoders. Stage 2 fine-tunes the model on an in-house 74-patient ICI-treated cohort, conditioning generation of post-treatment CTs on pre-treatment CTs and clinical variables through the 'cbi-Adapter', then extracts U-Net features for response classification and survival prediction. The paper reports 0.74±0.20 balanced accuracy (a 21.24% relative improvement over the second-best baseline) and a 0.57±0.12 c-index for survival prediction, along with qualitative observations about tumor size changes in generated images.

Significance. If the reported protocol is leakage-free, the work is a creative contribution: it combines anatomy-guided generative conditioning, clinical-variable conditioning, and transfer learning from public CT data to a small immunotherapy cohort. The direction is clinically relevant, and the use of a pretrained diffusion U-Net as a feature extractor is a falsifiable design choice that could benefit the community. The paper also provides a clear statement that code will be released, which would help reproducibility. However, the evidence is currently insufficient to establish the headline claims: the dataset is very small (19 responders), the variance is high, no significance tests are reported, and the training/validation split for Stage 2 is ambiguous. The contribution is significant if these issues are resolved, but the paper as written does not yet support the claimed gains.

major comments (4)

[§3, Dataset and Implementations] The manuscript never states whether the Stage 2 generative model \hat{D}_\theta, the cbi-Adapter, and the clinical conditioning are retrained from scratch inside each of the five folds or trained once on all 74 patients before the classification and survival heads are evaluated. The sentence 'The total number of patients used are 74 and the training is done in a 5-fold cross-validation setting' describes the downstream evaluation but does not clarify the Stage 2 training split. If the diffusion model was trained on all patients, then for each test patient the frozen U-Net feature extractor has already been trained on that patient's post-treatment CT as a generation target, and the 0.74 balanced accuracy and 0.57 c-index in Tables 1 and 2 would be inflated by leakage of the target scans. Please state explicitly which components are trained inside each fold, provide the fold splits or code, and report results with Stage 2 retrained independently in each fold.
[§3, Dataset and Implementations] The pre-treatment and post-treatment CT pairs are cropped to (96, 96, 16) with [3.0, 3.0, 0.5] spacing, but no registration or alignment step is described. If the two scans are not spatially aligned, the conditioning on z0 and the target z1 are not voxel-corresponding, so the model cannot faithfully learn the anatomical change induced by immunotherapy, and the extracted features may encode misalignment artifacts rather than true treatment response. Please specify the registration method used for the paired scans or provide evidence that the pairs are inherently aligned.
[Table 1] Several baseline rows are identical (ViT-B/16 and Swin-Base, and also DenseNet-121+CX), which suggests a copy-paste or reporting error and undermines confidence in the baseline numbers. More importantly, the proposed method's balanced accuracy is 0.74±0.20 over five folds, and with only 19 responders the uncertainty is large; no significance tests, confidence intervals, per-fold results, or AUC values are reported. The claimed 21.24% improvement cannot be distinguished from noise under these conditions. Please report per-fold results, paired significance tests, and additional discrimination metrics.
[§2.2, Survival Analysis and Table 2] The survival prediction component is described only as 'features fed to a survival head' with a hazard function H(t) = ∫ h(u) du, without specifying the loss function, censoring indicator, time-to-event definition, or evaluation protocol. The reported c-index advantage of 0.03 over the baseline is small and cannot be interpreted without these details. Please provide a full description of the survival model, including how censoring is handled and how the c-index is computed.

minor comments (6)

[§2.2, final paragraph] The sentence 'The The overall learning objective L_f inalis defines as' contains typos and should be rewritten as 'The overall learning objective is defined as'.
[§2.2, cbi-Adapter] The acronym 'cbi-Adapter' is never expanded; please define what 'cbi' stands for at first use.
[§2.2, Equation (2)] The notation for RoPE is inconsistent: the text defines f as a feature vector in R^{|D|}, but then treats it as having height and width positions (h, w); the definition should be clarified, especially because the clinical variables are tabular and do not have an obvious spatial position.
[§3, Qualitative Analysis] The claim that tumor size decreases in generated post-treatment CTs for responders is based on visual inspection; please provide quantitative tumor volume or RECIST measurements on the generated images to support this observation.
[§3, Qualitative Analysis] The radiologist review is mentioned without details on the number of cases reviewed, blinding to the response labels, or the specific criteria used; please add these details.
[Table 3] The ablation compares Stage 1, Stage 2, and the combined method, but it does not isolate the contribution of the cbi-Adapter from the contribution of clinical conditioning; an additional ablation with clinical conditioning but without the cbi-Adapter would clarify the adapter's role.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the response and survival labels enter only in the final prediction heads, not in the generative diffusion objective, and the self-citations are not load-bearing.

full rationale

The paper's Stage 2 generative objective is L_final = E[||epsilon - epsilon_theta(z_t, t, c_hat)||^2], with c_hat combining pre-treatment CT embeddings and clinical variables. The RECIST response and survival labels appear only when training the classification and survival heads on features extracted from the frozen U-Net; they are never used to fit the diffusion model itself. Thus the central prediction is not a renamed input or a fitted parameter: post-treatment CT synthesis is an intermediate surrogate, not the label. The hand-set weighting lambda=5e-2 and temperature tau=0.5 are arbitrary but not circular. Self-citations to the authors' prior work (refs [5,6]) are used only as general support that diffusion models are effective in medical image generation; they do not supply a uniqueness theorem, forbid alternatives, or carry the load of the main claim. The paper does not explicitly state whether the Stage 2 generative model is retrained inside each 5-fold split, and if it were trained on all 74 patients before splitting, test-patient post-treatment CTs could leak into the feature extractor and inflate results. That is a data-leakage and validity risk, not a circularity that reduces a prediction to an input by construction. No specific circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several hand-set hyperparameters (tau, lambda, RoPE base) and on domain assumptions about paired CT alignment and the informativeness of clinical variables. No new physical entities are introduced; the cbi-Adapter is a model component.

free parameters (3)

Temperature tau in contrastive loss = 0.5
Set by hand in Eq. (1); affects anatomy embedding alignment.
Clinical conditioning weight lambda = 5e-2
Set in the final loss after Eq. (2); controls the strength of clinical embeddings in generation, with no sensitivity analysis.
RoPE base b = 1e5
Standard choice from RoFormer, but still a hand-set hyperparameter that shapes position encoding.

assumptions (4)

standard math DDPM forward and reverse processes model image distributions with Gaussian noise in the standard way.
Section 2 Preliminary.
domain assumption Pre-treatment CT, clinical variables, and anatomy masks contain sufficient information to synthesize post-treatment CT.
Core generative assumption in Stage 2; not validated independently.
domain assumption Pre- and post-treatment CTs are from the same patients and are comparable after cropping to fixed size; no explicit registration is stated.
Section 3: CTs cropped to (96,96,16); no registration method given.
domain assumption The 5-fold cross-validation protocol does not leak test patients into the Stage 2 diffusion model training.
Section 3 says training is done in 5-fold CV but does not specify nested training of the generative model.

invented entities (1)

cbi-Adapter
purpose: Joint attention module to fuse clinical variables and pre-treatment image embeddings for conditioning the diffusion model.
This is an architectural component, not a natural entity; its utility is only demonstrated in this paper's in-house experiments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ImmunoDiff: A Diffusion Model for Immunotherapy Response Prediction in Lung Cancer." pith.science (2026). https://pith.science/paper/FCL5BQP5

@misc{pith2026250523675,
  author       = {Pith},
  title        = {Pith review of: ImmunoDiff: A Diffusion Model for Immunotherapy Response Prediction in Lung Cancer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCL5BQP5}},
  note         = {Machine review of arXiv:2505.23675}
}

read the original abstract

Accurately predicting immunotherapy response in Non-Small Cell Lung Cancer (NSCLC) remains a critical unmet need. Existing radiomics and deep learning-based predictive models rely primarily on pre-treatment imaging to predict categorical response outcomes, limiting their ability to capture the complex morphological and textural transformations induced by immunotherapy. This study introduces ImmunoDiff, an anatomy-aware diffusion model designed to synthesize post-treatment CT scans from baseline imaging while incorporating clinically relevant constraints. The proposed framework integrates anatomical priors, specifically lobar and vascular structures, to enhance fidelity in CT synthesis. Additionally, we introduce a novel cbi-Adapter, a conditioning module that ensures pairwise-consistent multimodal integration of imaging and clinical data embeddings, to refine the generative process. Additionally, a clinical variable conditioning mechanism is introduced, leveraging demographic data, blood-based biomarkers, and PD-L1 expression to refine the generative process. Evaluations on an in-house NSCLC cohort treated with immune checkpoint inhibitors demonstrate a 21.24% improvement in balanced accuracy for response prediction and a 0.03 increase in c-index for survival prediction. Code will be released soon.

Figures

Figures reproduced from arXiv: 2505.23675 by the authors.

**Figure 1.** ImmunoDiff pipeline.(A) Vessels and lobes masks are pre-trained in a contrastive manner. (B) A DDPM model is trained on a large cohort of NSCLC patient CT images. (C) A ControlNet is trained with the vessel and lobe as controls. (D) Clinical variables are used as controls in addition with pre-treatment CT images to generate post-treatment CT images. (E) cbi-Adapter architecture. (F) The trained diffusion model is u… view at source ↗

**Figure 2.** Qualitative results. (A) CT images generated w/o pre-training and with pretraining but w/o anatomy-guided training. (B) CT image generated using anatomyguided training. Also shown are the lobes and vessels used as controls. (C) Posttreatment CT generated by ImmunoDiff for responder and non-responder cases. (tumors shown in red box). erate clinically accurate and high-quality CT images. Instead, the integration o… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages

[1]

Nature communications5(1), 4006 (2014)

Aerts, H.J., et al.: Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature communications5(1), 4006 (2014)

work page 2014
[2]

Science advances (2022)

Alilou, M., et al.: A tumor vasculature–based imaging biomarker for predicting re- sponse and survival in patients with lung cancer treated with checkpoint inhibitors. Science advances (2022)

work page 2022
[3]

Medical physics (2011)

Armato III, S.G., et al.: The lung image database consortium (lidc) and image database resource initiative (idri). Medical physics (2011)

work page 2011
[4]

Scientific data5(1), 1–9 (2018)

Bakr, S., et al.: A radiogenomic dataset of non-small cell lung cancer. Scientific data5(1), 1–9 (2018)

work page 2018
[5]

arXiv preprint arXiv:2410.00307 (2024)

Bhattacharya, M., Singh, G., Jain, S., Prasanna, P.: Radgazegen: Radiomics and gaze-guided medical image generation using diffusion models. arXiv preprint arXiv:2410.00307 (2024)

arXiv 2024
[6]

In: MIDL (2024)

Bhattacharya, M., et al.: Gazediff: A radiologist visual attention guided diffusion model for zero-shot disease classification. In: MIDL (2024)

work page 2024
[7]

NEJM373(17), 1627–1639 (2015)

Borghaei, H., et al.: Nivolumab versus docetaxel in advanced nonsquamous non– small-cell lung cancer. NEJM373(17), 1627–1639 (2015)

work page 2015
[8]

arXiv preprint arXiv:2211.12737 (2022)

Chambon, P., et al.: Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737 (2022)

arXiv 2022

Show all 48 references

[9]

arXiv preprint arXiv:2010.11929 (2020)

Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. arXiv preprint arXiv:2010.11929 (2020)

2020 arXiv
[10]

NEJM378(22), 2078–2092 (2018)

Gandhi, L., et al.: Pembrolizumab plus chemotherapy in metastatic non–small-cell lung cancer. NEJM378(22), 2078–2092 (2018)

2018
[11]

In: CVPR (2016)

He, K., et al.: Deep residual learning for image recognition. In: CVPR (2016)

2016
[12]

NEJM378(22), 2093–2104 (2018)

Hellmann, M.D., et al.: Nivolumab plus ipilimumab in lung cancer with a high tumor mutational burden. NEJM378(22), 2093–2104 (2018)

2018
[13]

NeurIPS (2020)

Ho, J., et al.: Denoising diffusion probabilistic models. NeurIPS (2020)

2020
[14]

European radiology experimental (2020)

Hofmanninger, J., et al.: Automatic lung segmentation in routine imaging is pri- marily a data diversity problem, not a methodology problem. European radiology experimental (2020)

2020
[15]

In: CVPR

Huang, G., et al.: Densely connected convolutional networks. In: CVPR. pp. 4700– 4708 (2017)

2017
[16]

arXiv preprint arXiv:2302.09778 (2023)

Huang, L., et al.: Composer: Creative and controllable image synthesis with com- posable conditions. arXiv preprint arXiv:2302.09778 (2023)

2023 arXiv
[17]

Bioengineering10(11), 1258 (2023)

Hung,A.L.Y.,etal.:Med-cdiff:conditionalmedicalimagegenerationwithdiffusion models. Bioengineering10(11), 1258 (2023)

2023
[18]

Cancer immunology research (2020)

Khorrami, M., et al.: Changes in ct radiomic features associated with lymphocyte distribution predict overall survival and response to immunotherapy in non–small cell lung cancer. Cancer immunology research (2020)

2020
[19]

In: WACV

Kim, J., et al.: Adaptive latent diffusion model for 3d medical image to image translation. In: WACV. pp. 7604–7613 (2024)

2024
[20]

Cancers14(13), 3265 (2022) 10 Bhattacharya et al

Lei, X., et al.: Lobe-specific analysis of sublobar lung resection for nsclc patients with tumors 2 cm. Cancers14(13), 3265 (2022) 10 Bhattacharya et al

2022
[21]

In: ICCV (2023)

Li, A.C., et al.: Your diffusion model is secretly a zero-shot classifier. In: ICCV (2023)

2023
[22]

In: CVPR (2023)

Li, Y., et al.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)

2023
[23]

Cancer Imaging (2024)

Liao, C.Y., et al.: Personalized prediction of immunotherapy response in lung can- cer patients using advanced radiomics and deep learning. Cancer Imaging (2024)

2024
[24]

Theranostics13(15), 5223 (2023)

Lin, Q., et al.: Visualizing vasculature and its response to therapy in the tumor microenvironment. Theranostics13(15), 5223 (2023)

2023
[25]

In: ICCV (2021)

Liu, Z., et al.: Swin transformer: Hierarchical vision transformer using shifted win- dows. In: ICCV (2021)

2021
[26]

In: ICMLA (2022)

Mahmoud, A.Y., et al.: Machine learning experiments with artificially generated big data from small immunotherapy datasets. In: ICMLA (2022)

2022
[27]

Cancer Treatment and Research Communications31, 100540 (2022)

Mbanu, P., et al.: Clinico-pathological predictors of clinical complete response in rectal cancer. Cancer Treatment and Research Communications31, 100540 (2022)

2022
[28]

BMC Pulmonary Medicine24(1), 106 (2024)

Nilssen, Y., et al.: Distribution and characteristics of malignant tumours by lung lobe. BMC Pulmonary Medicine24(1), 106 (2024)

2024
[29]

In: MICCAI Workshop on Deep Generative Models

Pinaya, W.H., et al.: Brain imaging generation with latent diffusion models. In: MICCAI Workshop on Deep Generative Models. pp. 117–126. Springer (2022)

2022
[30]

arXiv preprint arXiv:2305.11147 (2023)

Qin, C., et al.: Unicontrol: A unified diffusion model for controllable visual gener- ation in the wild. arXiv preprint arXiv:2305.11147 (2023)

2023 arXiv
[31]

JAMA oncology (2024)

Rakaee, M., et al.: Deep learning model for predicting immunotherapy response in advanced non- small cell lung cancer. JAMA oncology (2024)

2024
[32]

Science348(6230), 124–128 (2015)

Rizvi, N.A., et al.: Mutational landscape determines sensitivity to pd-1 blockade in non–small cell lung cancer. Science348(6230), 124–128 (2015)

2015
[33]

In: CVPR

Rombach, R., et al.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684–10695 (2022)

2022
[34]

In: MICCAI

Ronneberger, O., et al.: U-net: Convolutional networks for biomedical image seg- mentation. In: MICCAI. pp. 234–241. Springer (2015)

2015
[35]

EBioMedicine86(2022)

She, Y., et al.: Deep learning for predicting major pathological response to neoad- juvant chemoimmunotherapy in non-small cell lung cancer: A multicentre study. EBioMedicine86(2022)

2022
[36]

Cancer Imaging24(1), 143 (2024)

Shrestha, B., et al.: Current trends in the characterization and monitoring of vas- cular response to cancer therapy. Cancer Imaging24(1), 143 (2024)

2024
[37]

CA: a cancer journal for clinicians

Siegel, R.L., et al.: Cancer statistics, 2024. CA: a cancer journal for clinicians

2024
[38]

Neurocomputing568, 127063 (2024)

Su, J., et al.: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing568, 127063 (2024)

2024
[39]

Phar- macy and Therapeutics42(8), 514 (2017)

Ventola, C.L.: Cancer immunotherapy, part 3: challenges and future trends. Phar- macy and Therapeutics42(8), 514 (2017)

2017
[40]

Radiotherapy and Oncology186, 109793 (2023)

Wang, X., et al.: Cancer immunotherapy response prediction from multi-modal clinical and image data using semi-supervised deep learning. Radiotherapy and Oncology186, 109793 (2023)

2023
[41]

Radiology: Artificial Intelligence5(5), e230024 (2023)

Wasserthal, J., et al.: Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence5(5), e230024 (2023)

2023
[42]

BJR| Arti- ficial Intelligence1(1), ubae013 (2024)

Webber, G., et al.: Diffusion models for medical image reconstruction. BJR| Arti- ficial Intelligence1(1), ubae013 (2024)

2024
[43]

In: MICCAI

Wolleb, J., et al.: Diffusion models for medical anomaly detection. In: MICCAI. pp. 35–45. Springer (2022)

2022
[44]

In: MICCAI

Xing, Z., et al.: Cross-conditioned diffusion model for medical image to image translation. In: MICCAI. pp. 201–211. Springer (2024)

2024
[45]

IEEE Transactions on Medical Imaging (2024) Title Suppressed Due to Excessive Length 11

Xu, Y., et al.: Medsyn: Text-guided anatomy-aware synthesis of high-fidelity 3d ct images. IEEE Transactions on Medical Imaging (2024) Title Suppressed Due to Excessive Length 11

2024
[46]

In: ICCV

Zhang, L., et al.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836–3847 (2023)

2023
[47]

IEEE Transactions on Medical Imaging (2024)

Zhang, Z., et al.: Diffboost: Enhancing medical image segmentation via text-guided diffusion model. IEEE Transactions on Medical Imaging (2024)

2024
[48]

NeurIPS36, 11127–11150 (2023)

Zhao,S.,etal.:Uni-controlnet:All-in-onecontroltotext-to-imagediffusionmodels. NeurIPS36, 11127–11150 (2023)

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Nature communications5(1), 4006 (2014)

Aerts, H.J., et al.: Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature communications5(1), 4006 (2014)

work page 2014

[2] [2]

Science advances (2022)

Alilou, M., et al.: A tumor vasculature–based imaging biomarker for predicting re- sponse and survival in patients with lung cancer treated with checkpoint inhibitors. Science advances (2022)

work page 2022

[3] [3]

Medical physics (2011)

Armato III, S.G., et al.: The lung image database consortium (lidc) and image database resource initiative (idri). Medical physics (2011)

work page 2011

[4] [4]

Scientific data5(1), 1–9 (2018)

Bakr, S., et al.: A radiogenomic dataset of non-small cell lung cancer. Scientific data5(1), 1–9 (2018)

work page 2018

[5] [5]

arXiv preprint arXiv:2410.00307 (2024)

Bhattacharya, M., Singh, G., Jain, S., Prasanna, P.: Radgazegen: Radiomics and gaze-guided medical image generation using diffusion models. arXiv preprint arXiv:2410.00307 (2024)

arXiv 2024

[6] [6]

In: MIDL (2024)

Bhattacharya, M., et al.: Gazediff: A radiologist visual attention guided diffusion model for zero-shot disease classification. In: MIDL (2024)

work page 2024

[7] [7]

NEJM373(17), 1627–1639 (2015)

Borghaei, H., et al.: Nivolumab versus docetaxel in advanced nonsquamous non– small-cell lung cancer. NEJM373(17), 1627–1639 (2015)

work page 2015

[8] [8]

arXiv preprint arXiv:2211.12737 (2022)

Chambon, P., et al.: Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737 (2022)

arXiv 2022