DADD Ulcerative Colitis Model Generates Severity Images, but It Is Not a Diagnostic Reader

DADD has generated ulcerative colitis endoscopy images across four severity grades, but the model does not diagnose patients from new colonoscopy images. The DADD ulcerative colitis model instead synthesizes controlled disease progression while trying to preserve the anatomy shown in a source image. That distinction matters because headlines about AI “reading” severity can imply a clinical function the published research does not establish.

Researchers Umut Dundar and Alptekin Temizel developed DADD at Middle East Technical University in Ankara, Turkey. The acronym stands for Disentangled Anatomy-Disease Diffusion. Their model separates structural information from disease-related texture, then generates an image representing a requested Mayo Endoscopic Score.

The peer-reviewed DADD study was published online on August 18, 2026, in the Journal of Medical Imaging. Its central result is not an autonomous diagnostic system. It is evidence that controllable synthetic images can improve a separate severity classifier when added to real training data.

That puts DADD into a different contest from systems such as UC-SCALE, which directly assign severity and location scores to colonoscopy footage. The core question is whether synthetic disease progression can solve a persistent data problem without introducing clinically misleading features.

What the DADD Ulcerative Colitis Model Actually Changed

DADD turns one endoscopic view into controlled synthetic versions of disease severity while attempting to keep the patient-specific structure consistent.

Ulcerative colitis causes inflammation in the lining of the colon. Clinicians use endoscopy to examine features such as redness, bleeding, vascular changes, erosions, and ulcers. These observations help determine disease activity and inform treatment decisions.

The Mayo Endoscopic Score, or MES, sorts endoscopic appearance into four ordered categories from 0 through 3. MES 0 represents inactive disease, while MES 3 represents severe inflammation. This ordinal structure gives the categories a meaningful direction, unlike four unrelated labels.

DADD uses that ordered scale as a control signal. A researcher can provide an endoscopy image and request a synthetic version associated with another severity level. The model then changes disease-related appearance while trying to retain the anatomy and composition of the source.

That is a harder task than generating a generic image labeled “severe.” Mucosal texture and anatomical structure occupy the same pixels. A conventional image encoder can therefore mix patient-specific anatomy with the signs of disease that the model is supposed to alter.

DADD addresses this problem with separate representations for anatomy and severity. Its anatomy pathway starts with a pretrained image encoder. A second, ordinal embedding represents the requested disease level.

The system also includes a feature purifier, which uses attention to suppress channels correlated with disease in the anatomical representation. The intended result is a cleaner structural signal that carries less information about the source image’s existing severity.

A triple-pathway cross-attention mechanism then supplies the diffusion network with purified anatomy, the target disease state, and their interaction. The network routes these inputs differently across its layers. Deeper features tend to influence global structure, while shallower features preserve finer textures.

The last major component is delta steering. This training-free control signal describes a direction between severity embeddings, such as movement from MES 1 toward MES 2. It lets the model make an explicit disease transition in a single generation pass.

The result is better understood as counterfactual visualization than diagnosis. DADD asks what an existing scene might look like under another severity condition. It does not establish what score a clinician should assign to an unseen patient.

That difference also explains the project’s evaluation strategy. The researchers measured whether the generated images resembled the requested grade, whether their visual distribution approached real data, and whether they helped train a downstream classifier.

On the Labeled Images for Ulcerative Colitis dataset, the hierarchical DADD version reached a regression-judge quadratic weighted kappa of 81.53 and a root-mean-square error of 0.46. Real images evaluated through the same framework produced 83.28 QWK and 0.47 RMSE.

Quadratic weighted kappa measures agreement while penalizing large ordinal errors more heavily than nearby mistakes. A prediction that confuses MES 0 with MES 3 therefore counts more seriously than confusion between adjacent grades.

The model also recorded a Fréchet inception distance of 27.20. FID compares statistical features from generated and real image sets, with lower values usually indicating greater similarity. It remains an indirect image metric rather than proof of clinical fidelity.

Most importantly, the addition of DADD images improved a downstream classifier. Real-only training produced 74.84% accuracy and 83.14 QWK. Adding synthetic images raised those measures to 77.75% and 85.77%, respectively.

The increase is modest but meaningful for the paper’s actual claim. Synthetic images helped rebalance training data and improved classification within the study’s experimental setting. The researchers did not show that DADD itself could replace endoscopists or clinical scoring systems.

Why Ulcerative Colitis AI Needs Better Training Data

The pressure comes from an uneven medical-image dataset, not from a shortage of model architectures.

Severe cases are typically less common than mild or inactive examples in labeled datasets. That imbalance can teach a classifier to favor common categories while struggling with rare but clinically important findings.

Medical datasets also face practical limits that internet-scale image collections do not. Colonoscopy images contain sensitive health information, require specialist annotation, and vary across patients, imaging equipment, clinical centers, and preparation quality.

An endoscopy frame can include blur, fluid, glare, instruments, or an incomplete view of the mucosa. The visible pattern can also vary within one patient’s colon. A single image therefore captures only part of a larger examination.

The public LIMUC dataset used by DADD contains 11,276 images from 564 patients. Its labels cover the four MES grades, but the distribution is uneven. That makes it useful for research while illustrating why rare-category augmentation remains attractive.

Traditional augmentation can crop, flip, rotate, or adjust the color of a real image. These operations increase visual variety, but they do not create a medically coherent transition from inactive to severe disease.

Generative models offer a more ambitious option. They can attempt to produce new examples of bleeding, ulcers, altered vascular patterns, or other severity-linked textures. Researchers can then add these examples to underrepresented classes.

However, naive generation creates a second problem. A model might change anatomy along with pathology, producing a different scene rather than a plausible progression of the source. It could also repeat artifacts or correlations found in the training set.

DADD is designed around this conflict. Its architecture treats anatomy as something to preserve and disease appearance as something to control. That gives its synthetic examples a clearer experimental purpose than generic class-conditioned generation.

The approach also uses the ordered relationship among MES grades. A progression from 0 to 1 differs from a jump between unrelated image classes. Delta steering tries to encode that directional change instead of asking the network to learn each category in isolation.

This distinction is important for AI developers. A synthetic dataset becomes more valuable when researchers know which factors the generator should preserve and which it should change. Without that separation, augmentation can quietly enlarge the wrong variation.

Yet a better class balance does not automatically produce a better clinical tool. Synthetic examples remain derived from the assumptions and biases of the original dataset. They cannot introduce patient populations, equipment characteristics, or disease manifestations absent from the source data.

DADD’s classifier improvement was measured on the same broader dataset environment used to develop and assess the method. The study did not report a prospective multicenter clinical trial or real-time use during colonoscopy.

That leaves the system under a familiar form of pressure. It must prove that controlled generation adds medically useful diversity, not merely images that match the scoring behavior of another machine-learning model.

The issue extends beyond DADD. A 2022 endoscopy classification study trained convolutional networks on HyperKvasir images and reported an 87.50% best accuracy for its selected task. That work also emphasized variability in endoscopic assessment and the preliminary nature of AI evidence.

The newer DADD research shifts attention upstream. Rather than introducing another standalone classifier, it targets the training data that classifiers consume. Its significance rests on whether that intervention transfers across institutions and model families.

Synthetic Progression Versus Direct Clinical Scoring

The primary contest is between generating controlled training examples and directly scoring real endoscopy footage.

Direct scoring systems answer the clinical question more plainly. They process real images or video, assign a severity estimate, and sometimes map the distribution of inflammation through the colon.

DADD serves a different layer of the workflow. It produces synthetic data and visualizations that other systems can use. Its output can support classifier development, but it is not presented as a clinical decision from a new examination.

This difference prevents a simple leaderboard comparison. A generator’s FID score and augmentation benefit do not answer the same question as a diagnostic system’s agreement with expert readers.

UC-SCALE provides a useful reference point. That system includes an image-quality filter, an MES scoring component, and a camera-localization algorithm. It was trained and tested with 4,326 sigmoidoscopy videos from phase 3 etrolizumab trials.

The published UC-SCALE evaluation reported a kappa of 0.80 against central readers at the colon-section level. It also found correlations between the system’s outputs and biomarkers, histology, clinician assessments, and patient-reported outcomes.

UC-SCALE therefore evaluates real examinations and connects its scores to several clinical measures. DADD evaluates whether generated images retain anatomy, represent requested severity, and strengthen a downstream classifier.

Another 2026 system, AI-ESe, analyzes complete endoscopic videos through multiple stages. It checks image quality, detects pauses by the endoscope, estimates severity, and creates an inflammatory heatmap across the examination.

The AI-ESe results included a quadratic weighted kappa of 0.80 for severity assessment. Its pause-detection component reduced mean absolute temporal disagreement in position estimates from 39 seconds to 23 seconds.

AI-ESe also exposed a weakness in worst-lesion scoring. Among videos assigned the highest conventional endoscopy subscore, the proportion of moderately to severely inflamed mucosa ranged from 17.9% to 100%.

That range means two examinations can share the same maximum grade while carrying very different total inflammatory burdens. Systems that analyze extent attempt to preserve information lost by a single worst-area score.

DADD does not solve the extent problem. It operates on endoscopic images and produces controlled visual transitions. Still, its anatomy-preserving design addresses a related need: generating disease changes without discarding the identity of the observed scene.

The routes can eventually complement each other. A synthetic progression model might produce rare examples for training a direct scoring system. A video-level model might then evaluate real footage using representations strengthened by those examples.

That combination remains hypothetical until independent studies test it. DADD’s reported classifier improvement does not establish that it will help UC-SCALE, AI-ESe, or another commercial system.

The difference between generation and scoring also shapes risk. A direct scoring error can influence the interpretation of a real patient’s examination. A generator introduces risk earlier by altering the data used to teach later systems.

Upstream errors can be harder to detect. If synthetic images contain a recurring artifact associated with MES 3, a classifier might learn that artifact instead of clinically meaningful inflammation.

The same problem can occur when a generated feature looks plausible to a general image metric but appears wrong to an inflammatory bowel disease specialist. FID cannot determine whether a vascular pattern, ulcer margin, or bleeding appearance is clinically coherent.

DADD’s regression judge offers an ordinal assessment, but it is still a model-based evaluator. Agreement between a generator and another algorithm can overstate quality when both systems inherit similar dataset biases.

For medical AI teams, the practical lesson is precise. The DADD ulcerative colitis model is a data-generation method with encouraging downstream evidence. It should not be described as a validated reader of patient severity.

What the Numbers Do Not Establish

DADD’s gains support further testing, but they do not establish diagnostic accuracy, patient benefit, or safe clinical deployment.

The study’s strongest practical number is the downstream accuracy increase from 74.84% to 77.75%. That represents a gain of 2.91 percentage points after synthetic data entered classifier training.

Its QWK gain, from 83.14 to 85.77, suggests that the augmented classifier also made better severity-ordered predictions. This metric is valuable because MES mistakes differ in clinical distance.

Even so, one dataset and one experimental pipeline cannot settle generalization. A model can improve on a familiar distribution while failing on images from another hospital, endoscope manufacturer, patient population, or annotation protocol.

The LIMUC data came from a single institutional setting in Turkey. DADD’s authors are also based in Turkey, although the public dataset lets other researchers reproduce and challenge the work.

External validation would need deliberately different data. Useful tests would include varied endoscope systems, bowel preparation quality, demographic groups, and inflammatory patterns. Video should also matter because clinical endoscopy is not a collection of independent still images.

The model’s framing as longitudinal synthesis needs similar care. DADD creates images representing different severity levels while preserving aspects of a source image. It does not follow a patient through repeated examinations and predict that individual’s future disease course.

A generated healing image is therefore a conditional visual hypothesis. It is not evidence that a treatment will produce that change, nor a forecast of when healing will occur.

The MES control also narrows what the model represents. MES is widely used, but it compresses complex visible findings into four categories. The underlying mucosa can vary substantially within each grade.

Established scoring research has documented reader disagreement. Work that developed the Ulcerative Colitis Endoscopic Index of Severity found weighted kappa values ranging from 0.30 to 0.65 across evaluated descriptors and reader comparisons.

That uncertainty affects the labels used to train generative and discriminative systems. A model cannot cleanly separate anatomy and disease when the underlying definition of severity contains observer variation.

Synthetic data can also amplify label assumptions. If the training images associate a specific visual feature with one grade, the generator can reproduce that connection even when specialists would interpret it differently in another context.

The feature purifier introduces another open question. It aims to remove disease-correlated information from anatomy tokens. Researchers still need to determine whether it also removes structural details that legitimately interact with disease appearance.

Conversely, some disease information might remain in the supposedly purified representation. That leakage could make severity control less independent than intended, especially for transitions far from the source grade.

The paper’s reported similarity between generated and real regression-judge results is encouraging. Yet the two RMSE values, 0.46 for generated images and 0.47 for real images, do not mean the synthetic images are clinically equivalent to real examinations.

The generator also needs expert review focused on failure modes. Specialists should inspect whether ulcers, vascular changes, bleeding, and texture transitions remain plausible at local and global levels.

A reader study could compare DADD images with real images while asking clinicians to identify the severity, detect implausible features, and rate diagnostic usefulness. Agreement alone would not be enough if errors cluster around treatment-relevant boundaries.

Researchers should also test whether classifiers trained with DADD augmentation improve on external real data. That is the clearest next step because the augmentation claim is central to the work.

A prospective study would go further by fixing the model before evaluation and measuring its effect on predefined clinical tasks. Such a design would reduce the opportunity to tune the system around the test environment.

Regulatory questions remain premature but unavoidable. Software used only for research augmentation faces a different risk profile from software that supplies patient-specific diagnostic information.

If DADD-generated images eventually become part of a regulated training pipeline, developers will need records describing their provenance and role. They will also need controls preventing synthetic cases from entering clinical records as if they were observed images.

None of these gaps invalidates the reported result. They define its boundary. DADD shows that anatomy-aware diffusion can create useful experimental data, not that synthetic endoscopy is ready to guide care.

Three Signals That Will Decide Whether DADD Matters

External classifier gains, specialist review, and video-level testing will determine whether DADD becomes useful infrastructure or remains a promising benchmark result.

The first signal is performance on an external real-image dataset. Researchers should train comparable classifiers with and without DADD images, then freeze both systems before testing them elsewhere.

A repeatable improvement would strengthen the claim that DADD adds useful disease variation. A gain that disappears outside LIMUC would suggest the generator mainly reinforces its source distribution.

The most informative analysis would separate results by MES grade. Synthetic augmentation should help the rare categories it is meant to rebalance without reducing performance on common categories.

Calibration also matters. A classifier can gain accuracy while becoming too confident about errors. Medical applications need probability estimates that remain trustworthy across populations and equipment changes.

The second signal is blinded review by inflammatory bowel disease specialists. Reviewers should assess local pathology, anatomical preservation, transition realism, and the presence of generation artifacts.

This study should include difficult adjacent transitions, such as MES 1 to MES 2. Those boundaries can carry more ambiguity than a visually dramatic shift from inactive to severe disease.

Specialists should also compare DADD with simpler augmentation and other generative approaches. A complex architecture earns its place only if its anatomy-disease separation produces a measurable advantage.

If reviewers consistently detect impossible or misleading features, the classifier gain becomes less persuasive. It could indicate that downstream models benefit from shortcuts unavailable or inappropriate in clinical images.

The third signal is expansion from still images to full video and spatial disease extent. Colonoscopy unfolds over time, and clinicians interpret consecutive views rather than isolated frames.

Video introduces motion, repeated mucosal views, blur, changing illumination, and variable examination speed. A generator that works on still images can produce temporal flicker or contradictory pathology when applied frame by frame.

DADD would also need a way to preserve coherent anatomy across a sequence. Each generated frame should remain consistent with neighboring frames and the requested disease state.

Direct-scoring projects show why this matters. AI-ESe explicitly detects endoscope stalling to reduce oversampling, while UC-SCALE estimates position within the colon. Both treat video structure as part of the measurement problem.

A successful video extension would strengthen DADD’s relevance to real development pipelines. Failure to maintain temporal consistency would keep its role limited to still-image experiments.

These tests should not be replaced by another favorable image-generation metric. The central promise is that disentangling anatomy from disease yields better medical training data. Validation must therefore measure the performance, realism, and transportability of those data.

For developers, DADD offers a useful architecture pattern. Separate the factors that should remain fixed from those the model should change, then test whether that separation improves a downstream task.

For clinicians, the immediate message is narrower. The system does not provide a new patient-facing severity score, treatment recommendation, or prognosis.

For patients, generated images should not be interpreted as predictions of personal disease progression. They represent modeled appearances attached to a four-level scale.

The DADD ulcerative colitis model deserves attention because it targets a real bottleneck in medical AI: scarce, unevenly distributed expert-labeled images. Its early results show a measurable classifier benefit and a clearer form of severity control.

The decisive evidence must now come from real data the model did not already know. Will DADD augmentation improve externally tested classifiers, survive blinded specialist review, and remain coherent across video? Those three results will determine whether synthetic progression becomes dependable medical-imaging infrastructure.