List of Papers Browse by Subject Areas Author List
Abstract
In this paper, we propose CT2-DiT, a simple yet powerful framework for abdominal non-contrast CT (NCCT) to contrast-enhanced CT (CECT) translation, with a particular focus on focal liver lesions (FLLs). CECT plays an important role in the clinical diagnosis of FLLs by providing a clear contrast between normal liver tissue and lesions. However, due to its invasive nature and potential risks, neural network-based translation from NCCT to CECT has become an attractive research direction. Diffusion Transformers (DiT) exhibit strong modeling capacity and effectively capture complex image structures, making them particularly promising for modeling abdominal CT images with complex liver anatomy. But their application to CECT synthesis is often limited by the need for large-scale training data, which is difficult to obtain in practice. To address this issue, we construct a medium-scale CT dataset using publicly available unlabeled data for pre-training, and improve generalization through manually synthesized CT-specific features and a spatial-intensity decoupled data augmentation strategy. We further enhance performance by incorporating deep supervision and improving data efficiency. With only 774 paired CT slices, CT2-DiT generates realistic and diagnostically relevant CECT images, outperforming existing methods across multiple evaluation metrics and demonstrating the previously underexplored potential of DiT in CECT synthesis. Notably, CT2-DiT trained on only 50% of the data still outperforms all baselines trained on the full dataset in structural fidelity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1524_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
LiTS dataset: https://www.kaggle.com/datasets/harshwardhanbhangale/lits-dataset
TCGA-LIHC dataset: https://www.cancerimagingarchive.net/collection/tcga-lihc/
RAOS dataset: https://github.com/Luoxd1996/RAOS
HCC-TACE-Seg dataset: https://www.cancerimagingarchive.net/collection/hcc-tace-seg/
BibTex
@InProceedings{ZhuHan_CT2DiT_MICCAI2026,
author = { Zhu, Han AND Song, Yu AND Nishikawa, Ikuko},
title = { { CT2-DiT: Enabling Abdominal Non-contrast to Contrast-Enhanced CT Translation with Diffusion Transformers in the Low-Data Regime } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This article presents CT2-DiT, a DiT-based framework for abdominal NCCT-to-CECT translation. To address the limited availability of paired medical data and the high data demand of DiT models, the authors combine pretraining on unlabeled public CT data with several training strategies, including CT-specific feature injection and spatial-intensity decoupled augmentation, to improve synthesis performance under small-sample conditions. The method is evaluated on 774 paired CT slices and compared against GAN-based and U-Net-based diffusion methods. In addition to image synthesis results, the authors also conduct downstream lesion classification experiments to provide a preliminary assessment of whether the synthesized images retain task-relevant information.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper explores the use of a pure DiT model for NCCT-to-CECT synthesis, which has some exploratory value from a methodological perspective. 2.The study presents a relatively complete training pipeline to address data scarcity, including pretraining, augmentation, and supervision strategies.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The proposed model has already leveraged information from around 250,000 CT slices during the first-stage pretraining, whereas other comparative methods have not undergone similar pretraining. In addition, the existing ablation results do not clearly demonstrate the actual performance of this method under a no-pretraining scenario. Based on the current results, one can only roughly estimate numerically that the method may be inferior to BBDM without pretraining. Therefore, the current experiments provide insufficient support for the intrinsic advantages of the proposed method, and the benefits brought by pretraining may have a significant impact on the final performance. 2.The manuscript mentions generating pseudo-class labels through K-Means on image features, but it remains unclear what the class label 𝑙 actually represents and whether it is defined consistently across the first-stage pretraining and second-stage fine-tuning. From the current description, one can only infer that pseudo-class labels are obtained by clustering image features, while the actual semantics of these pseudo-classes remain unclear. If this conditional label is related to lesion categories or downstream task labels, it may introduce a risk of label leakage; if not, the authors should further clarify how this condition is defined and used across the two training stages. 3.The use of a “Train on Synthetic, Test on Real” protocol to assess the claimed clinical utility is understandable, especially given the limited amount of real data. However, this experiment seems more reflective of the surrogate training value of synthetic data, and it is less clear whether it is sufficient to support the stronger claim of “clinical utility. ” 4.The proposed Spatial-Intensity Decoupled Data Augmentation, as currently described, does not seem to constitute decoupling in a strict methodological sense, but rather the use of two common augmentation strategies with different application scopes. This design does not seem substantially different from conventional augmentation practices in paired medical imaging tasks. Therefore, the authors should clarify in what sense this strategy is genuinely novel compared with existing augmentation methods; otherwise, it seems insufficient to present it as a methodological innovation. 5.The motivation of the proposed CT-specific feature injection is understandable; however, based on Algorithm 1, the method appears more like a binary selection map constructed from CT intensity ranges and empirically chosen thresholds, rather than a true “texture map” in a strict sense. The current method does not seem to fully support the manuscript’s claims about highlighting soft tissue boundaries and providing finer-grained guidance. The authors should further clarify the actual advantage of this binary map over other prior information, such as organ or lesion masks; otherwise, the method currently appears more like an empirically designed structural cue than a fully convincing innovative approach.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I give this paper a weak reject because, although the problem is meaningful and the paper makes a reasonable attempt to introduce a pure DiT framework for NCCT-to-CECT synthesis under data-limited settings, the overall evidence is still not sufficiently convincing. Several claimed innovations appear closer to training strategies or empirical design choices, and their methodological novelty is not yet clearly established. In addition, some key design choices remain insufficiently explained, such as the semantics of the pseudo-class labels and the consistency of the class label across the two training stages. The current experiments also do not clearly separate the benefit of pretraining from the contribution of the proposed method itself, and the downstream evaluation provides only limited support for the stronger claim of clinical utility. Therefore, while the work shows some potential, the current version still leaves room for further strengthening in terms of novelty positioning, methodological clarity, and experimental support.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed my previous concerns, particularly regarding the fairness of baseline comparisons, the clarification of label usage, and the refinement of clinical claims. The additional explanations on the pre-training strategy, data augmentation design, and downstream evaluation further strengthen the motivation and technical soundness of the work. Overall, the paper presents a well-justified and practically meaningful approach. I recommend acceptance.
Review #2
- Please describe the contribution of the paper
This work introduces the first application of a pure Diffusion Transformer (DiT) for NCCT-to-CECT translation. It specifically addresses the “data-hungry” nature of Transformers by proposing a three-pronged strategy: (1) Pre-training on unlabeled CT slices, (2) injecting binarized CT texture maps to recover high-frequency details lost in VAE compression, and (3) a spatial-intensity decoupled augmentation strategy.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Addressing Data Scarcity: Successfully training a DiT on only 774 paired slices is a significant achievement. The “Half-data” experiments (Table 1) show remarkable robustness. Pre-training Strategy: The use of MedDINO for deep supervision during pre-training allows the model to learn semantically meaningful medical features before fine-tuning. Feature Injection: The “NCCT Binarization” (Algorithm 1) is a simple yet effective way to provide structural priors, effectively acting as a “ControlNet” for medical textures. Clinical Relevance: The downstream classification task (92.91% accuracy) demonstrates that the synthesized images are not just visually pleasing but contain lesion-specific pathological features (e.g., FNH hyper-enhancement).
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
VAE Limitations: The authors acknowledge that the pre-trained VAE blurs fine details. While they attempt to mitigate this with texture injection, a comparison with a pixel-space DiT (like JiT: https://github.com/LTH14/JiT) would be beneficial. Intensity Normalization: The binarization threshold (t=0.05≈50HU) is very specific to the liver. It is unclear how this would generalize to other organs without manual tuning. 2D-Only: The method operates on individual 2D CT slices and does not enforce inter-slice consistency. This may cause discontinuous artifacts, inconsistent lesion appearance, and anatomical incoherence in volumetric CT data, which is critical for real clinical reading.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The transition from U-Nets to Transformers is a major trend in vision, and this paper provides a roadmap for doing so in medical translation tasks where data is scarce. The technical strategies (MedDINO supervision and texture injection) are well-justified and yield impressive results on a difficult task. However, exploring 3D real-world data would benifit the work significantly.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed most of my concerns in the rebuttal. Considering that the overall quality of this paper meets the acceptance criteria of MICCAI, I am willing to recommend acceptance.
Review #3
- Please describe the contribution of the paper
The paper introduces CT2-DiT, a generative framework that adapts the Diffusion Transformer (DiT) architecture for the translation of abdominal non-contrast CT (NCCT) scans into contrast-enhanced CT (CECT) images. The primary contribution is a data-centric methodology designed to mitigate the scarcity of paired medical datasets. This approach includes large-scale pre-training on 250,000 unlabeled public CT slices to learn general abdominal representations and a spatial-intensity decoupled data augmentation strategy to improve generalization. To preserve high-frequency anatomical details, the authors introduce a CT-specific feature injection strategy using binarized texture maps. Furthermore, the framework incorporates deep supervision by aligning intermediate DiT features with representations from a pre-trained medical vision foundation model, MedDINO v3.The performance of the proposed method is validated on a private dataset of focal liver lesions, where it demonstrates superior structural fidelity compared to established GAN and U-Net-based diffusion baselines despite being trained on a limited set of 774 paired slices. Finally, the clinical utility of the synthesized images is assessed through a downstream lesion classification task.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The framework demonstrates exceptional data efficiency by achieving superior structural fidelity compared to established GANs and U-Net-based diffusion models even when trained on a limited dataset of only 774 paired slices.
2.The paper provides a strong clinical validation through a downstream lesion classification task using a train on synthetic, test on real protocol. The model achieved a high accuracy of 92.91 percent in distinguishing between four types of focal liver lesions, which significantly outperformed the others SOTA models.
3.The model provides positive qualitative performance for preserving lesion-specific pathology.
4.Ablation study confirms the contribution of each component. Starting from a DiT baseline, performance gains were observed sequentially by adding large-scale pre-training, binarized texture maps, spatial-intensity decoupled augmentation, and IREPA-based supervision, with each module providing measurable improvements in intensity-based metrics.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.According to the ablation study in Table 3, the most substantial performance gains are driven by pre-training and data augmentation. These are well-established techniques in the literature. The more specific contributions, such as the Binarize Map and iREPA supervision, provide comparatively marginal incremental improvements.
2.The authors compare their method against six competing models. However, it is not clarified if these baselines were granted the same advantages of extensive pre-training and decoupled data augmentation. If not, the comparison is scientifically invalid as it measures the impact of the training regime rather than the architectural superiority of CT2-DiT.
3.The model operates strictly on 2D axial slices. This is a major limitation in 3D medical imaging, as it neglects volumetric context and inter-slice consistency. Modern SOTA typically uses 3D patch-based or 2.5D approaches. Please add a discussion around this limitation and potential future works.
4.Resizing CT data to 256 x 256 slices risks losing fine-grained diagnostic details. Given that foundation models like MedDINO are designed for high-resolution dense features, this downscale is counter-intuitive and appears to be a concession to memory constraints. Please add explanations on this preprocessing choice.
5.While 86 patients is a reasonable starting point, the resulting 774 paired slices averages to just ~9 slices per patient. This confirms a highly selective, slice-based approach that ignores full-volume reality. Also, critical technical parameters such as voxel size, interpolation methods used during resizing, and specific registration protocols for aligning multi-phase scans are not adequately defined.
6.Table 2 lacks an NCCT-only baseline, which is essential to quantify the diagnostic value added by synthesis. Furthermore, the NCCT+CECT upper bound (96.07%) mentioned in the text should be included in the table to ease readibility.
7.The claims around the preservation of anatomical structures is unjustified. Downstream segmentation performance of the synthesized CECT would be required to validate your findings.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper is a well-executed technical report showing that Diffusion Transformers can achieve competitive results in data-scarce medical environments when supported by heavy pre-training and augmentation. However, the scientific contribution is limited by the lack of 3D/volumetric context and the reliance on a small, private 2D slice dataset. The authors should have discussed the limitations of this approach and explored potential 2.5D implementations to better capture inter-slice anatomical continuity. Also, the perceived SOTA performance likely comes down to the extensive pre-training rather than a fundamental framework breakthrough. Finally, the paper provides insufficient detail regarding the dataset, voxel spacing, registration techniques, preprocessing pipelines, and slice selection criteria. These omissions make it difficult for the research community to replicate the findings or assess the impact of data quality on model performance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their rebuttal, but several critical methodological concerns remain unsatisfactorily addressed.
- The justification for not granting baselines the same pre-training and data augmentation advantages is unconvincing. Attributing superior performance to the architecture while comparing it against models denied the same extensive data pipeline invalidates the fairness of the SOTA benchmark.
- Defending a purely 2D slice-based approach as a clinical choice to avoid z-axis variation bypasses the reality of modern 3D medical imaging pipelines. A discussion around 2.5D implementation must be discussed if 3D-patch-based approach is not feasible.
- The authors argue that global metrics mask the localized improvements of their core contributions (Binarize Map and iREPA). While qualitative sharpness is noted, the numerical gains of these proposed modules remain marginal compared to standard pre-training and augmentation, meaning the claimed architectural breakthrough is not robustly reflected in the quantitative data. Consequently, while the downstream classification task shows promise, the paper ultimately presents an extensively pre-trained 2D engineering pipeline rather than a validated methodological advancement for 3D clinical workflows.
Author Feedback
Baseline Comparisons (R1, R4) Differences stem from architecture designs, not protocols: Unpaired Pre-training (PT): Baselines require paired data or meaningful labels. Our DiT naturally digests 250K unpaired unlabeled dataset via noise-to-image PT. U-Net baselines require redesigns to digest unpaired data. This is a fundamental architectural capability, not an unfair advantage. Decoupled DA as a DiT Recipe: CNN/U-Nets have strong spatial inductive biases; DiTs do not. Our DA is heavily engineered for DiTs to avoid intensity memorization. Applying this DiT-specific DA to baselines (e.g., GANs) disrupts training dynamics (ref: Training GANs with Limited Data). For fairness, we evaluated baselines using their official, optimal settings rather than artificially impairing them with mismatched DA. Performance Without DA: Even without DA (Tab. 3, “+ Binarize Map”), our model still performs better on lesion regions; due to space, we omit qualitative ablations. It achieves 0.8160 SSIM, outperforming PPT (0.7840) and SynDiff (0.8094). 2D Operations (R2, R4) Our 2D approach is a choice driven by clinical and data constraints: Z-axis Variations: Scans have varying slice thicknesses. 2D operations on max-lesion slices avoid interpolation artifacts and focus capacity on pathology. 3D Overfitting: 86 paired volumes are insufficient to train a 3D generative Transformer without severe overfitting. PT Compatibility: Our 250K PT dataset pools multi-center data (some only available as 2D slices). A 2D architecture is strictly necessary to leverage this diverse, large-scale data. This establishes a solid baseline. We will add these justifications and discuss 3D extensions as future work. R1 Label Leakage: Our PT K-Means labels capture dataset domains (e.g., scanners). Fine-tuning labels only indicate contrast phases (ART/PV). Neither uses lesion categories; thus, no label leakage exists. Map vs. Masks: Manual masks suffer from imprecise boundaries and empirically yielded no gains. Our binarization directly thresholds pixels to extract precise edges, successfully enhancing sharpness and textures. Terminology: We agree. True “clinical utility” requires radiologist studies. We will revise this to state that our downstream task demonstrates the preservation of pathological features and surrogate training value. R2 HU Generalization: 50 HU is a clinical constant for the liver, not a tuned hyperparameter. To generalize to other organs, one simply applies their established clinical HU values instead of manual tuning. Pixel-space DiT: We agree on VAE limits. However, our tests with a JiT showed worse performance. Pixel-space attention likely requires massive data to converge on medical textures, confirming VAEs remain necessary for data-scarce regimes. We will discuss this. R4 Metrics Limitations: Global metrics (SSIM/PSNR) are dominated by the background, where DA helps most. Lesions occupy very few pixels, so modules targeting them (map and iREPA) show small global numerical gains. Qualitatively, however, they are essential for recovering fine lesion boundaries and textures that DA misses. Dataset: We used 774 slices to focus strictly on lesion-bearing slices, as healthy slices lack diagnostic value here. Instead of volumetric registration, slices were clinically paired by maximum tumor cross-section, ensuring exact pathological correspondence for our 2D framework. Class. over Seg.: Clinically, lesion localization is usually pre-determined; the challenge is FLL classification based on contrast patterns. We chose classification to best measure the preservation of these critical textures. Table 2: We will add the NCCT-only baseline (81.86%) and the real NCCT+CECT upper bound to Table 2 to explicitly show our synthesis boosts accuracy. MedDINO Upsampling: We bridge our DiT and high-res MedDINO features using an asymmetric alignment layer (See Sec. 2.1). This architecture maintains memory efficiency without sacrificing MedDINO’s high-fidelity supervision.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper was reviewed by three experts in the field and received the following recommendations: two Weak Rejects and one Weak Accept. While the reviewers noted several strengths, including the promising performance of the proposed framework in addressing data scarcity and its clinical value, they also raised several concerns, including a lack of inter-slice consistency, potentially unfair comparisons with baseline methods, and insufficiently justified contributions of the proposed modules. The authors are encouraged to address these concerns in the rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After the rebuttal, two reviewers recommended acceptance, while one reviewer maintained a Reject recommendation. While the paper has some limitations, including the need for further clarification of the baseline comparisons, it also demonstrates several strengths, including promising performance and a well-motivated framework. Therefore, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Reviewer 4 raised concerns regarding limitations in the evaluation. However, the novelty of the proposed method is sufficient, and the comparison results with previous methods are favorable. Therefore, I believe this paper is worthy of acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
There are 2 Accepts and 1 Reject after rebuttal. Reviewers agreed that the paper addresses an important problem and presents a promising strategy for adapting DiTs to low-data medical image translation. The rebuttal clarified several key concerns, including the role of pretraining, the pseudo-label design, the motivation behind the augmentation strategy, and the scope of the clinical claims. While concerns remain regarding the fairness of baseline comparisons and the limitations of the 2D setting, the overall contribution is technically sound and supported by the experimental results. The presentation could also be improved, including figures, tables, and overall organization. I recommend acceptance.
