Technical Summary: Disentangled Anatomy-Disease Diffusion (DADD) for Controllable Ulcerative Colitis Progression Synthesis

1. Problem Statement

The synthesis of longitudinal medical images at controllable disease stages is hindered by the entanglement of patient-specific anatomical features and pathological textures. In the context of Ulcerative Colitis (UC), disease severity follows a continuous ordinal progression defined by the Mayo Endoscopic Score (MES), ranging from 0 (normal) to 3 (severe). Existing datasets, such as LIMUC, suffer from severe class imbalance and a lack of longitudinal follow-up due to the invasive nature of repeated endoscopies.

Current generative approaches face a critical "entanglement bottleneck." Methods that condition generation on reference images (e.g., IP-Adapter using CLIP encoders) inevitably embed the source disease pathology into the anatomical representation. When attempting to synthesize a counterfactual progression (e.g., changing a source image from MES 3 to MES 0), the model receives conflicting signals: it must preserve the anatomical structure while simultaneously erasing the source pathology and generating a new severity level. This semantic conflict often leads to structural artifacts or the failure to remove existing disease features.

2. Methodology

The proposed framework, Disentangled Anatomy-Disease Diffusion (DADD), addresses this by explicitly separating anatomical identity from disease severity within a latent diffusion model. The architecture integrates three novel components into a standard Stable Diffusion backbone:

2.1. Additive Ordinal Embedder (AOE)

To model the continuous nature of MES, the framework employs an AOE that represents severity as a cumulative sum of learnable increments. A higher MES level includes all features of lower levels plus additional pathology. This allows for the generation of arbitrary intermediate severity levels via continuous interpolation.

2.2. Feature Purifier

To resolve the entanglement between anatomy and source disease, a Feature Purifier is introduced. This module operates in three stages:

1. Identification: A cross-attention layer queries the image tokens against the source disease embedding (derived from the source MES label) to identify disease-correlated channels.
2. Gating: A gating MLP generates a sigmoid mask to control the suppression of these identified features.
3. Erasure: A gated residual subtraction removes the disease component from the image tokens, yielding "purified" anatomy tokens ($e_{clean}$) that retain structural identity but lack source pathology.

2.3. Triple-Pathway Cross-Attention

Instead of concatenating all conditioning tokens, DADD utilizes a split-injection mechanism with three distinct pathways injected into the U-Net:

• Anatomy Pathway: Processes the purified anatomy tokens ($e_{clean}$) using pretrained projections to preserve structural reconstruction.
• Disease Pathway: Processes the target disease tokens ($e_{aoe}^t$) using separate, bias-free projections.
• Delta Pathway: Processes the directional severity change ($\Delta$) derived from the difference between target and source embeddings.

Frequency-Aware Routing: The framework employs resolution-dependent routing gates ($g_a, g_d$) to weight the contribution of anatomy versus disease at different U-Net depths. Low-resolution blocks (capturing global structure) are weighted heavily toward anatomy, while high-resolution blocks (capturing fine texture) are weighted toward disease. This ensures that disease edits are confined to fine-grained pathological textures without disrupting global mucosal structure.

2.4. Delta Steering

To enable precise control over disease transitions without the computational cost of Classifier-Free Guidance (CFG), the authors introduce Delta Steering. This is a training-free, single-pass directional signal derived from the projected difference between target and source ordinal embeddings.

• Mechanism: $\Delta = \text{proj}(E[y_t]) - \text{proj}(E[y_s])$.
• Operation: During inference, $\Delta$ is processed through the bias-free disease projections to inject a signed, magnitude-proportional signal into the attention mechanism.
• Benefit: This eliminates the need for unconditional forward passes, reducing inference latency by approximately 2× compared to standard CFG.

3. Key Contributions

• Disentanglement via Erasure: A cross-attention-based Feature Purifier that dynamically identifies and suppresses disease-correlated channels from image tokens, effectively disentangling anatomy from pathology.
• Frequency-Aware Injection: A Triple-Pathway Cross-Attention mechanism with fixed routing gates that directs anatomical and pathological signals to appropriate U-Net layers based on resolution hierarchy.
• Delta Steering: A novel, single-pass control mechanism for monotonic disease progression that provides signed, magnitude-proportional control, offering a more efficient alternative to CFG.
• Targeted Augmentation: An empirical demonstration that synthesizing subject-specific longitudinal data effectively rebalances skewed class distributions, enhancing downstream classification performance.

4. Experimental Results

The framework was validated on the LIMUC dataset (10,284 images) using Stable Diffusion v1.4.

• Generation Quality: While some variants (DADD-I, DADD-U) achieved lower Fréchet Inception Distance (FID) scores, they failed to produce discriminative pathological features, resulting in near-random performance in downstream tasks. The Hierarchical Routing (DADD-H) configuration, which aligns disease injection with high-resolution layers, achieved the best balance.
• Downstream Classification: DADD-H significantly improved classification accuracy and Quadratic Weighted Kappa (QWK) compared to baselines. In a "Synthetic Only" training scenario (no real data), DADD-H achieved 61.68% accuracy and 70.74 QWK, outperforming the IP-AOE baseline (54.54% / 42.20 QWK).
• Severity Fidelity: A Regression Judge trained on real data evaluated the synthetic outputs. DADD-H achieved 81.53 QWK and 0.46 RMSE, closely matching the performance of models trained on real data (83.28 QWK / 0.47 RMSE). This confirms the model genuinely transforms pathology rather than replicating the source.
• Inference Speed: DADD variants achieved roughly a 2× speedup over CFG-based baselines (e.g., 187.3 ms/img vs. 314.6 ms/img at batch size 8) due to the single-pass delta steering mechanism.
• Qualitative Analysis: Visual comparisons showed that DADD-H successfully erased existing pathological textures when synthesizing lower severity levels (e.g., MES 3 $\to$ 0), whereas IP-Adapter baselines failed to remove source disease, resulting in entangled generations.

5. Significance and Claims

The paper claims that DADD provides a robust framework for synthesizing longitudinal medical images where patient-specific anatomy is preserved while disease severity is controllably modulated. The authors emphasize that:

1. Distributional similarity (FID) is insufficient for validating conditional progression synthesis; models with low FID may fail to alter pathology meaningfully.
2. Disentanglement is critical: Explicitly erasing source disease information from the anatomical embedding is necessary to avoid semantic conflicts during counterfactual generation.
3. Efficiency: The delta steering mechanism offers a computationally efficient alternative to CFG, making controllable synthesis more practical for clinical applications.
4. Clinical Utility: The ability to generate balanced, subject-specific longitudinal data directly addresses the endemic class imbalance in medical datasets, improving the training of downstream diagnostic classifiers.

The authors note that performance remains sensitive to the steering scale ($\lambda_{steer}$), with values $\ge 5$ introducing structural artifacts, and suggest future work on per-class steering schedules and extension to other longitudinal medical modalities.