Gated Conditional Diffusion Model (GCDM)

• Gated Conditional Diffusion Model (GCDM) is a latent diffusion framework designed for generating realistic mammograms with controlled lesion synthesis and global anatomical coherence.
• It employs a soft-mask concatenation branch to preserve overall breast anatomy and a gated lesion-conditioning branch to fuse radiomic and geometric features dynamically.
• Quantitative results demonstrate improved performance (FID: 26.77, Mass IoU: 86.30) compared to baseline models, highlighting enhanced image realism and fine-grained lesion control.

Gated Conditional Diffusion Model (GCDM) denotes a latent denoising diffusion framework for controllable mammogram synthesis that is designed to jointly synthesize holistic mammogram images and localized lesions. In the formulation introduced in "Joint Holistic and Lesion Controllable Mammogram Synthesis via Gated Conditional Diffusion Model" (Li et al., 25 Jul 2025), the model combines a soft-mask concatenation branch, which represents breast, lesion, and their transitional regions, with a gated conditioning branch that dynamically selects and fuses radiomic and geometric lesion properties. The architecture is therefore organized around two coupled objectives: preservation of global breast anatomy and fine-grained control of small lesion regions.

1. Problem setting and design objective

GCDM addresses controllable mammogram synthesis on the VinDr-Mammo dataset using only CC-view images. The target conditioning variables are a breast mask describing the global breast shape, a lesion mask describing lesion location and extent, and lesion-specific attributes drawn from radiomic and geometric descriptors. The paper motivates this setup by arguing that prior diffusion-based mammogram synthesis methods, although more stable than GANs, mostly provide global control via masks or segmentation guidance and often fail to model the fine interaction between lesion and tissue. The reported failure modes include abrupt lesion boundaries, lesions that do not match the intended size or shape, poor consistency between lesion mask and generated appearance, and limited diversity in lesion-specific characteristics (Li et al., 25 Jul 2025).

Within that framing, GCDM is not an unconditional image generator. It is a conditional mammogram generator that attempts to satisfy two requirements simultaneously: the synthesized image must remain globally anatomical and realistic, while also allowing explicit control over lesion shape, texture, and lesion–tissue blending. This suggests that the model is structured to treat local lesion realism as inseparable from whole-image anatomical coherence.

2. Latent diffusion formulation

GCDM is built on a latent denoising diffusion model, in the style of Stable Diffusion. A clean mammogram image $x_0 \in \mathbb{R}^{3 \times H \times W}$ is first encoded by a pretrained VAE encoder $\mathcal{E}$ into a latent representation

$$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$

Gaussian noise is then added according to the standard forward diffusion process,

$$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$

where $t$ is the diffusion timestep, $\beta_t$ defines the noise schedule, and $\bar{\alpha}_t = \prod_{s=1}^{t}(1-\beta_s)$. The denoiser $\epsilon_\theta$ is trained to predict the injected noise from the noisy latent together with the conditioning inputs. At inference, a random latent $z_T \sim \mathcal{N}(0,I)$ is iteratively denoised to $\hat{z}_0$ and decoded back to image space as $\mathcal{E}$0 (Li et al., 25 Jul 2025).

The backbone choice matters because lesion control is implemented in latent space rather than directly in pixel space. A plausible implication is that the authors adopt latent diffusion not only for computational convenience, but also to make global structure and local lesion constraints compatible within a single denoising process.

3. Soft-mask concatenation for anatomical coherence

The first conditioning pathway is a soft-mask concatenation branch. The paper uses a three-channel binary mask

$\mathcal{E}$1

whose channels correspond to background, breast tissue, and lesion or mass. To improve lesion–tissue continuity, the lesion channel is blurred with a Gaussian operator $\mathcal{E}$2, producing a soft mask. The semantic interpretation is explicit: the breast region captures global breast anatomy and tissue envelope, the lesion region specifies the mass area to be synthesized, and the transitional region introduced by Gaussian blurring models the boundary zone through which lesions blend into surrounding tissue (Li et al., 25 Jul 2025).

Conditioning is imposed by channel-wise concatenation of the noisy latent and the encoded soft mask. The training objective is

$\mathcal{E}$3

This branch provides what the paper describes as holistic control. It does not attempt to specify lesion appearance in a fine-grained manner; rather, it anchors lesion placement and breast anatomy in a common denoising trajectory. The use of a blurred lesion label is especially important because it avoids treating the lesion as a pasted hard-boundary object.

4. Gated lesion-conditioning branch

The second conditioning pathway is the gated-fusion-based Mass Control Branch. Its purpose is to provide lesion-specific control beyond what mask concatenation alone can express. The branch uses two descriptor families. First, radiomic features $\mathcal{E}$4 are extracted with PyRadiomics from the lesion mask $\mathcal{E}$5 and the corresponding image $\mathcal{E}$6. The paper states that $\mathcal{E}$7 includes shape: 9 dimensions, histogram: 18 dimensions, GLSZM: 16 dimensions, and GLCM: 24 dimensions. If no lesion exists, the vector is set to zero, $\mathcal{E}$8. Second, the lesion mask $\mathcal{E}$9 is embedded with CLIP and passed through convolutional layers to obtain geometric candidate features $$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$0 (Li et al., 25 Jul 2025).

The paper argues that directly concatenating radiomic and geometric features can be suboptimal because the two feature types live in very different spaces, some combinations are redundant, others may be incompatible, and not every radiomic–geometric pairing is clinically meaningful. Gating is introduced as the mechanism that evaluates candidate combinations and keeps only the most relevant ones. The cross-combination tensor is

$$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$1

The gate score is then defined as

$$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$2

where $$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$3 captures intra-feature statistics and $$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$4 models inter-feature interactions. The final fused control condition is

$$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$5

The resulting vector $$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$6 is injected into the diffusion UNet via cross-attention, yielding the full objective

$$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$7

This is the architectural core of GCDM in the strict sense of the paper: the model is “gated conditional” because lesion conditions are not passed to the denoiser wholesale, but are selectively reweighted and filtered before cross-attention-based control.

5. Optimization, data, and evaluation protocol

The reported implementation uses PyTorch with Stable Diffusion v1.5 as the backbone, AdamW as the optimizer, learning rate $$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$8, a noise schedule with $$z_0 = \mathcal{E}(x_0) \in \mathbb{R}^{4 \times 32 \times 32}.$$9 linearly from 8.5e-4 to 0.012, 400 training epochs, batch size 32 per GPU, 2 × NVIDIA A100 80GB, diffusion steps $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$0, and 50 inference steps. Classifier-free guidance is implemented by randomly masking conditioning with probability 0.1 during training and using guidance scale 7.5 at inference. The gate-module hyperparameters are $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$1, and the Gaussian blur uses $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$2 (Li et al., 25 Jul 2025).

Experiments are conducted on VinDr-Mammo using only CC images. The split is train 6704, validation 748, and test 1871, with mass-containing images distributed as train 373, validation 45, and test 104. Preprocessing removes black background using the max connected component, resizes images to $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$3, and applies truncation normalization for contrast enhancement. For mass images, bounding-box annotations are fed to MedSAM to obtain instance masks, radiomic features are extracted with PyRadiomics, and features are Min-Max normalized using training statistics. Baselines are SPADE, pSp, SR3, ControlNet, and Seg-Diff. The evaluation metrics are FID for realism and distribution match, Mass IoU for lesion controllability, Breast IoU for global anatomical alignment, and PA for pixel-wise consistency between generated mask and target mask (Li et al., 25 Jul 2025).

The protocol makes the intended evaluation axis explicit: realism is not treated as sufficient, and controllability is not reduced to mask overlap alone. Instead, the benchmark combines global anatomical agreement and lesion-level precision.

6. Quantitative performance and ablation structure

On the main quantitative comparison, GCDM reports FID 26.77, Mass IoU 86.30, Breast IoU 97.63, and PA 98.41. The corresponding baseline values are SPADE: 51.60, 83.57, 97.67, 98.40; pSp: 45.92, 71.51, 95.08, 96.65; SR3: 39.28, 84.02, 97.08, 98.04; ControlNet: 33.67, 79.36, 97.42, 98.27; and Seg-Diff: 30.50, 79.34, 96.94, 98.01. The paper further reports that, compared to the second-best result, GCDM improves FID by 12.2% and Mass IoU by 2.71% with statistical significance (Li et al., 25 Jul 2025).

The ablation studies isolate the contribution of lesion conditioning and gating. The baseline using only mask plus noisy latent yields FID 31.25 and Mass IoU 83.48. Adding the Lesion Control Branch gives 29.60 and 84.97. Adding radiomics features gives 28.21 and 85.78. Adding gated fusion gives 26.77 and 86.30. The soft-mask ablation compares hard mask, which yields FID 27.23 and Mass IoU 86.19, with soft $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$4, which yields 28.84 and 86.83, and soft $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$5, which yields 26.77 and 86.30. The paper also states that larger $$z_t = \sqrt{\bar{\alpha}_t}\,\mathcal{E}(x_0) + \sqrt{1-\bar{\alpha}_t}\,\epsilon, \quad \epsilon \sim \mathcal{N}(0,I),$$6 keeps improving FID while Mass IoU drops. This indicates an explicit trade-off: stronger blur improves image realism, while excessive blur weakens lesion boundary control (Li et al., 25 Jul 2025).

Qualitatively, the paper reports that SPADE shows lower tissue realism, pSp shows weak lesion control, SR3 struggles with multiple lesions, and ControlNet and Seg-Diff often produce lesions inconsistent with the input mask. GCDM is described as producing mammograms closer to real images, with richer breast tissue detail and smoother lesion–tissue transitions. The appendix also reports a downstream classification experiment in which augmenting training data with GCDM-generated samples improves performance for both ResNet-50 and ViT-B/16. This suggests that the model’s utility is not confined to visual plausibility alone.

7. Scope, limitations, and terminological ambiguity

The paper implies several limitations. The method depends on good lesion masks and radiomic extraction quality; very large mask blurring improves realism but reduces precise lesion control; the evaluation is primarily on CC-view mammograms from one dataset; and the lesion-conditioning branch relies on extracted features rather than fully end-to-end semantic lesion understanding (Li et al., 25 Jul 2025). These constraints are relevant because they delimit what “controllability” means in the current formulation: the model is controllable with respect to supplied masks and handcrafted lesion descriptors, not with respect to a broader clinical semantic ontology.

The acronym GCDM is also not unique in arXiv usage. "Gated Condition Injection without Multimodal Attention: Towards Controllable Linear-Attention Transformers" (Liu et al., 29 Mar 2026) uses the name GCDM/GateControl for a different controllable diffusion framework tailored to linear-attention transformer backbones such as SANA, where gating is token-wise fusion after the linear-attention layer and the target application is general controllable visual generation rather than mammogram synthesis. "Guaranteed Conditional Diffusion: 3D Block-based Models for Scientific Data Compression" (Lee et al., 18 Feb 2025) does not explicitly define a model called Gated Conditional Diffusion Model at all; its actual names are Guaranteed Conditional Diffusion with Tensor Correction and Guaranteed Conditional Diffusion. "Conditional Diffusion with Less Explicit Guidance via Model Predictive Control" (Shen et al., 2022) studies sparse explicit guidance in diffusion sampling and describes a gating-like schedule over timesteps, but it does not introduce a model named GCDM. In encyclopedia usage, therefore, GCDM most precisely denotes the mammogram synthesis framework of (Li et al., 25 Jul 2025), while other occurrences of the acronym or related phrasing refer to distinct conditional diffusion designs.