DADD: Disentangled Anatomy-Disease Diffusion

• The paper introduces a novel conditional diffusion model (DADD) that disentangles anatomy and disease to enable controllable ulcerative colitis progression synthesis.
• It uses purified image tokens with ordinal disease embeddings and triple-pathway cross-attention to accurately modulate Mayo Endoscopic Scores.
• Delta Steering provides a single-pass, inference-time control signal, achieving precise disease editing with competitive performance on the LIMUC dataset.

Searching arXiv for the named DADD paper and closely related anatomy–disease/anatomy-guided diffusion work to ground the article. Disentangled Anatomy-Disease Diffusion (DADD) denotes a conditional latent diffusion framework in which patient-specific anatomy and disease severity are represented through separate conditioning pathways so that disease progression can be edited while preserving structural identity. In its explicit formulation for ulcerative colitis (UC) endoscopy, DADD is designed to synthesize images at requested Mayo Endoscopic Score (MES) stages from a real source image, using purified anatomy tokens, ordinal disease tokens, and an inference-time directional control signal termed Delta Steering (Dundar et al., 3 May 2026).

1. Clinical problem and conceptual basis

DADD is motivated by controllable progression synthesis in UC endoscopy, where repeated endoscopy is invasive and dense longitudinal image sequences for the same patient are scarce. The operational objective is to generate a new endoscopic image that preserves the source patient’s anatomy while changing only disease severity to a requested target stage. The clinical grading variable is the Mayo Endoscopic Score, an ordinal four-level scale in which MES 0 denotes normal mucosa with visible vascular pattern, MES 1 mild erythema with decreased vascular pattern, MES 2 marked erythema with friability and erosions, and MES 3 spontaneous bleeding with ulceration (Dundar et al., 3 May 2026).

The core difficulty is that anatomy and disease are visually entangled in the same frame. Patient-specific anatomy includes mucosal folds, lumen geometry, tissue layout, and structural texture, whereas disease severity appears through erythema, vascular pattern loss, friability, erosions, bleeding, and ulceration. If a reference image is used as a conditioning signal, its embedding preserves not only anatomy but also the source disease state. The DADD formulation therefore treats the raw image embedding as an entangled representation, summarized in the paper as

$\mathbf{e}_{\mathrm{img} \approx A \oplus D_s,$

with $A$ denoting anatomy and $D_s$ source disease (Dundar et al., 3 May 2026).

This problem belongs to a broader medical-imaging line of work on anatomy–pathology separation. Earlier work on brain MRI disentanglement argued that clinically plausible factorization should not enforce full independence when anatomy and pathology are statistically dependent, but should still make them separately controllable (Hu et al., 2022). DADD adopts that general orientation, but implements it in a diffusion framework with explicit anatomy and disease pathways.

2. Anatomy and disease representations

The DADD architecture is built on Stable Diffusion v1.4 and uses a frozen VAE encoder-decoder together with a frozen CLIP ViT-L/14 image encoder. The anatomy branch begins by extracting CLIP image features from the source endoscopic image $\mathbf{I}$, then compressing them with a trainable Perceiver Resampler into $N=16$ tokens of dimension $D=768$:

$\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$

These tokens are intended to preserve patient-specific visual structure, but they also leak source pathology and therefore cannot be used directly as a pure anatomy prompt (Dundar et al., 3 May 2026).

The disease branch uses a separately trained Additive Ordinal Embedder (AOE), which encodes MES as a cumulative ordinal variable rather than an unrelated categorical class. For $K=4$ severity levels, the embedding is

$$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$

where $\mathbf{b} \in \mathbb{R}^D$ is a learnable healthy base vector and $A$0 are learnable severity increments. The target MES embedding is projected into $A$1 disease tokens,

$A$2

again with $A$3 and $A$4 (Dundar et al., 3 May 2026).

This ordinal construction reflects the clinical assumption that higher MES stages accumulate lower-stage abnormalities plus additional pathology. The same representation also supports continuous interpolation:

$A$5

so the disease-control space is not restricted to integer MES labels (Dundar et al., 3 May 2026).

In DADD, the anatomy and disease factors are therefore not separate latent variables inferred symmetrically from the input. Instead, anatomy is supplied by a purified image-conditioned token set, while disease is supplied by an ordinal target embedding. The disentanglement problem is delegated to a dedicated purification module and to resolution-aware conditioning inside the denoising network.

3. Feature Purifier and Triple-Pathway Cross-Attention

The Feature Purifier is the mechanism that converts entangled image tokens into anatomy-dominant tokens. It uses the source disease embedding $A$6 to identify disease-correlated content inside $A$7. First, a disease-related component is extracted by cross-attention:

$A$8

A gating MLP then predicts a suppression mask

$A$9

and the cleaned anatomy tokens are produced by gated residual subtraction:

$D_s$0

The purifier is thus a cross-attention-based erasure mechanism acting in token-feature space rather than on pixels (Dundar et al., 3 May 2026).

Condition injection into the U-Net is handled by Triple-Pathway Cross-Attention. Instead of concatenating all conditions into a single attention pool, DADD uses three separate pathways: an anatomy pathway from $D_s$1, a disease pathway from $D_s$2, and a delta pathway used only at inference. Given U-Net hidden states $D_s$3, the shared query is $D_s$4. The anatomy and disease attention outputs are

$D_s$5

$D_s$6

A crucial architectural distinction is that the anatomy pathway uses Stable Diffusion’s pretrained key/value projections with bias, while the disease pathway uses separate bias-free projections warm-started from pretrained weights (Dundar et al., 3 May 2026).

The three pathways are fused with resolution-dependent routing:

$D_s$7

This routing exploits the U-Net hierarchy. In the best-performing hierarchical configuration, low-resolution blocks emphasize anatomy with $D_s$8 and $D_s$9, while high-resolution blocks emphasize disease with $\mathbf{I}$0 and $\mathbf{I}$1. An inverted routing and a uniform routing were also evaluated and performed worse on severity fidelity (Dundar et al., 3 May 2026).

4. Delta Steering, training, and inference

Delta Steering is an inference-time directional control signal derived from the ordinal disease embeddings. It is defined as the post-projection difference between target and source severity tokens:

$\mathbf{I}$2

The corresponding attention output is

$\mathbf{I}$3

Because the subtraction is performed after projection and the disease-pathway projections are bias-free, a zero delta yields exactly zero pathway output. During training, the framework sets $\mathbf{I}$4, hence $\mathbf{I}$5 and $\mathbf{I}$6; the delta pathway is therefore exactly inactive during optimization (Dundar et al., 3 May 2026).

The diffusion model is trained in latent space with the standard $\mathbf{I}$7-prediction objective from DDPM, together with Min-SNR-$\mathbf{I}$8 weighting with $\mathbf{I}$9. The paper states this objective but does not explicitly reproduce the full forward diffusion equations. Inference uses DDIM with 50 steps (Dundar et al., 3 May 2026).

The trainable components are the U-Net, Perceiver Resampler, AOE, and Feature Purifier. The VAE encoder-decoder and CLIP image encoder are frozen. Training is joint rather than staged. Optimization uses AdamW with base learning rate $N=16$0, a $N=16$1 learning-rate multiplier for the Resampler and Purifier, cosine annealing to $N=16$2, two warmup epochs, FP16, gradient checkpointing, and EMA with $N=16$3. The diffusion schedule uses 1000 timesteps with linear $N=16$4, $N=16$5, and $N=16$6 (Dundar et al., 3 May 2026).

A notable point is that DADD does not introduce an explicit disentanglement loss, ordinal consistency loss, or purifier-specific regularizer. The separation of anatomy and disease is induced structurally through distinct embeddings, the purifier, and routed cross-attention, while the optimization signal remains the diffusion denoising objective (Dundar et al., 3 May 2026).

5. Empirical evaluation on LIMUC

The framework is validated on LIMUC, a UC endoscopy dataset that initially contains 11,276 labeled images from 564 patients. After removing 992 artifact images, the retained set contains 10,284 images. Images are cropped to $N=16$7 and resized to $N=16$8. The patient-level split is 7,180 training images, 1,661 validation images, and 1,443 test images. The class distribution is strongly imbalanced: MES 0 accounts for 54.4%, MES 1 for 27.0%, MES 2 for 11.0%, and MES 3 for 7.6% (Dundar et al., 3 May 2026).

Generation is evaluated against IP-Text and IP-AOE baselines using FID, CMMD, Improved Precision, and Improved Recall. On these distributional metrics, DADD-H does not dominate all baselines: IP-AOE obtains FID $N=16$9, whereas DADD-H obtains FID $D=768$0. However, the paper emphasizes that lower FID does not imply better controllable synthesis. DADD-I and DADD-U achieve still lower FID values, $D=768$1 and $D=768$2, yet under-modulate disease and remain too close to the conditioning image (Dundar et al., 3 May 2026).

This point becomes explicit in disease-control evaluation. A ResNet-18 regression judge, trained only on real data, predicts MES from synthetic images. The real-data reference reaches QWK $D=768$3 and RMSE $D=768$4. DADD-H reaches QWK $D=768$5 and RMSE $D=768$6, nearly matching that ceiling, whereas IP-AOE reaches QWK $D=768$7 and RMSE $D=768$8, DADD-I QWK $D=768$9 and RMSE $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$0, and DADD-U QWK $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$1 and RMSE $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$2. This is the strongest quantitative evidence that hierarchical routing yields correct target-severity realization rather than merely realistic-looking images (Dundar et al., 3 May 2026).

Downstream classification further separates realism from utility. When augmented with synthetic data, a ResNet-18 classifier improves over real-only training, which yields accuracy $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$3 and QWK $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$4. The mixed-training performance of DADD-H is accuracy $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$5 and QWK $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$6. The stricter synthetic-only setting shows a larger separation: DADD-H reaches accuracy $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$7 and QWK $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$8, versus IP-AOE at $\mathbf{e}_{\mathrm{img} \in \mathbb{R}^{N \times D}.$9 and $K=4$0, DADD-I at $K=4$1 and $K=4$2, and DADD-U at $K=4$3 and $K=4$4 (Dundar et al., 3 May 2026).

Delta Steering also changes the computational profile. With DDIM-50, FP16, and an RTX 6000 Ada, DADD-H requires $K=4$5 ms per image at batch size 8, corresponding to $K=4$6 images/s, whereas IP-Text requires $K=4$7 ms per image and $K=4$8 images/s, and IP-AOE requires $K=4$9 ms per image and $$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$0 images/s. The single-pass directional steering therefore yields roughly a $$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$1 speedup over CFG-based baselines (Dundar et al., 3 May 2026).

6. Relation to adjacent methods and limitations

DADD occupies a distinct position within recent medical generative modeling. Several closely related methods separate anatomy from another factor, but not anatomy from disease in the explicit DADD sense. In 4D cardiac MRI synthesis, anatomy-guided residual motion diffusion factorizes static anatomy and temporal motion while representing diagnosis only as conditioning, not as an independently disentangled disease factor (Cao et al., 25 Jun 2026). In weakly supervised brain anomaly reconstruction, conditional latent diffusion with pseudo-pathology preserves anatomy functionally while removing lesions, but disease remains a conditioning signal rather than a separate latent representation (Aguila et al., 11 Jun 2025). PathoSyn is closer to explicit anatomy–pathology separation because it models pathology as an additive residual on a stable anatomical substrate, but it operates through deviation-space diffusion and lesion support rather than through purified anatomy tokens and ordinal disease embeddings (Wang et al., 29 Dec 2025). Earlier structured variational models for brain MRI likewise argued that clinically plausible anatomy–pathology disentanglement should preserve dependency structure rather than enforce total independence (Hu et al., 2022).

A common misconception is that any anatomy-guided diffusion model constitutes DADD. The literature does not support that equivalence. Anatomy-conditioned normative modeling (Zhang et al., 2024), anatomy-aware disease-trajectory prediction (Yang et al., 15 May 2025), and anatomy-guided augmentation (Cao et al., 25 Jun 2026) all reduce anatomical confounding, but they do not necessarily provide an explicit anatomy branch, a disease branch, and a mechanism for editing disease while preserving anatomy. DADD, in its explicit UC formulation, is defined by precisely that separation (Dundar et al., 3 May 2026).

The framework also has clear limitations. It assumes that the source MES $$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$2 is known at inference, because both the Feature Purifier and Delta Steering depend on it. Performance is sensitive to the steering scale: $$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$3 under-modulates disease, $$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$4 provides the best balance, and $$E[k] = \mathbf{b} + \sum_{i=0}^{k-1} \boldsymbol{\delta}_i, \quad k \in \{0, \ldots, K{-}1\},$$5 introduces artifacts and anatomy degradation. Validation is confined to UC endoscopy on LIMUC, so cross-organ, cross-modality, and cross-disease generalization remain open. The framework also does not use a direct disentanglement loss or a reader-study-based clinical validation. A plausible implication is that the anatomy–disease decomposition remains structural and empirical rather than formally identifiable (Dundar et al., 3 May 2026).

Within the current literature, DADD is therefore best understood as an explicit, diffusion-based anatomy–disease conditioning framework rather than a generic label for anatomy-aware medical generation. Its defining contributions are the purification of image-derived anatomy tokens, ordinal disease embeddings that respect cumulative progression, resolution-aware routing inside the denoiser, and a training-free directional control signal that converts disease editing from a two-pass guidance problem into a single-pass conditional operation (Dundar et al., 3 May 2026).