Semantic-Curvature Mask Generator

Updated 6 July 2026

Semantic-Curvature Mask Generator is a conditional mask synthesis system that combines semantic inputs with spatial constraints to generate binary tumor masks.
It employs a differentiable curvature loss to enforce realistic boundary morphology, differentiating benign and malignant tumor shapes.
The framework integrates a hierarchical GAN architecture and explicit geometric regularization, enhancing downstream image generation, classification, and segmentation.

Searching arXiv for the primary paper and closely related mask-generation work. arxiv_search(query="(Pan et al., 10 Jul 2025)", max_results=5, sort_by="submittedDate") Searching arXiv for related papers on semantic mask generation, diffusion-conditioned masks, and geometry-aware mask synthesis. arxiv_search: {"query": "(Pan et al., 10 Jul 2025)", "max_results": 5, "sort_by": "submittedDate"} A Semantic-Curvature Mask Generator denotes a class of mask synthesis or mask-editing mechanisms that combine semantic conditioning with explicit or implicit geometric regularization in order to produce structurally plausible masks for downstream generation, editing, or recognition. In the narrowest and most explicit sense, the term refers to the Semantic-Curvature Mask Generator (SCMG) introduced for breast ultrasound tumor synthesis, where a conditional generator produces binary tumor masks under direct control of tumor type and bounding rectangle, and uses a differentiable curvature prior to enforce clinically realistic benign-versus-malignant boundary statistics (Pan et al., 10 Jul 2025). In a broader, interpretive sense suggested by adjacent work, the same label can cover systems in which semantic prompts, labels, or cross-modal signals are coupled to topology-aware, hull-based, edge-aware, or shape-aware mask generation, even when curvature is not modeled explicitly (Qian et al., 2024, Wang et al., 2024, Couairon et al., 2022).

1. Definition and Conceptual Scope

In the BUS framework of "Breast Ultrasound Tumor Generation via Mask Generator and Text-Guided Network:A Clinically Controllable Framework with Downstream Evaluation" (Pan et al., 10 Jul 2025), SCMG is defined as “a shape-aware module that synthesizes tumor masks using three clinical priors: tumor type (benign or malignant) for semantic control, bounding rectangle for spatial localization, and curvature regularization for boundary plausibility.” Formally, it is a conditional generator

$\hat{M} = G(z, k, s),$

where $z$ is latent noise, $k$ is tumor type, and $s$ is a bounding rectangle map. The output $\hat{M}$ is a binary tumor mask used as structural control for the image generator.

Two aspects are central to the term. The “semantic” component refers to conditioning on clinically or semantically meaningful variables such as tumor type, category labels, or text prompts. The “curvature” component refers to an attempt to regulate boundary shape so that masks match class-specific morphology rather than merely occupying the correct region. In SCMG, this curvature component is literal and differentiable. In other systems, however, the same idea is approximated through topology preservation, convex hull regularization, morphological opening, or score-difference localization rather than direct curvature penalties (Qian et al., 2024, Xue et al., 2023, Couairon et al., 2022).

A recurring ambiguity is terminological. Not every semantically controlled mask generator is curvature-aware in the strict differential-geometric sense. Conversely, not every boundary-regularized mask generator is semantically expressive. The distinctive contribution of SCMG is to join both properties in a single front-end mask generator whose output is intended for a downstream diffusion pipeline (Pan et al., 10 Jul 2025).

2. Position in Conditional Generative Pipelines

The canonical SCMG instantiation appears in a clinically controllable BUS synthesis pipeline with two forms of conditioning: structural guidance from tumor masks and semantic guidance from BI-RADS-aligned text. A VAE first encodes real BUS images into a latent space. A latent diffusion model then generates latent images conditioned on a ControlNet mask $m_0$ and a CLIP text embedding $\mathcal{C}_\theta(\tau_0)$ , after which the VAE decoder reconstructs the BUS image. At inference, because no real mask is available, SCMG supplies the mask $\hat{M}$ , making the system controllable and annotation-free at inference (Pan et al., 10 Jul 2025).

This front-end role has close analogues in other domains. In MaskFactory, mask generation is likewise separated from image generation: a mask editing stage first creates synthetic masks through rigid and non-rigid transformations, and a multi-conditional diffusion stage then produces paired image-mask outputs conditioned on the edited mask, its Canny edge map, and a category prompt (Qian et al., 2024). In face synthesis, a class-wise VAE with bi-directional LSTMs manipulates semantic parts at the mask level before a semantic image synthesis model such as SEAN produces the final RGB output (Fontanini et al., 2023). DiffEdit uses a different mechanism but preserves the same division of labor: the diffusion model first derives an automatic semantic mask by contrasting denoising predictions under different prompts, and that mask is then used to constrain the actual editing trajectory in latent space (Couairon et al., 2022).

This architecture pattern suggests a general division between layout control and appearance realization. A plausible implication is that Semantic-Curvature Mask Generators are most useful when a system benefits from separating the geometry of “where and what shape” from the appearance of “how it looks,” especially in settings where manual masks are unavailable or undesirable.

3. Architecture of the Explicit SCMG Formulation

SCMG in the BUS system adopts a conditional GAN-style design with a hierarchical SPADE-conditioned generator and a dual-branch discriminator (Pan et al., 10 Jul 2025). The generator begins by projecting latent noise $z \in \mathbb{R}^n$ through a fully connected layer into an initial feature map $F_0 \in \mathbb{R}^{C_0 \times H_0 \times W_0}$ . Tumor type $z$ 0 is embedded, spatially broadcast, and concatenated with $z$ 1, ensuring that semantic type information is present from the first stage onward. The bounding rectangle $z$ 2 is resized per scale and injected through SPADE normalization,

$z$ 3

so that each upsampling block is spatially constrained by the intended lesion location and size.

Each hierarchical block follows the sequence nearest-neighbor upsampling, SPADE with instance normalization, ReLU, and a $z$ 4 convolution. A final $z$ 5 convolution followed by a sigmoid yields a one-channel probability map in $z$ 6, which is thresholded to form the binary mask. The paper describes the output as having the same spatial resolution as the BUS image, for example $z$ 7 (Pan et al., 10 Jul 2025).

The discriminator contains a spatial branch and a frequency branch. The spatial branch evaluates shape, continuity, and topology using convolutional layers with instance normalization. The frequency branch computes the Fourier magnitude spectrum of the mask and processes it with convolutional layers to assess sharpness and fine boundary detail. The discriminator outputs both a real/fake probability and a tumor type prediction. This construction is notable because semantic supervision is not only injected into the generator through class embeddings, but also recovered from mask geometry through discriminator-side classification.

Related mask generators distribute these roles differently. MaskFactory allocates semantic editing to a text-guided non-rigid editor with mutual attention and allocates structural consistency to a discriminator operating on structural graphs of mask boundaries (Qian et al., 2024). XMask3D uses a frozen diffusion UNet as a 2D mask generator, conditions it on global 3D features, and attaches a Mask2Former-style head to produce class-agnostic masks and mask embeddings for later 3D-2D-text alignment (Wang et al., 2024). These variations preserve the same principle: mask generation is treated as an intermediate representation learning problem rather than a trivial preprocessing step.

4. Curvature Modeling, Semantic Control, and Boundary Statistics

What distinguishes SCMG from generic semantic mask generators is its explicit curvature formulation. For a binary mask $z$ 8, the paper extracts boundary pixels $z$ 9 and computes discrete first and second derivatives. The mean absolute curvature is then defined as

$k$ 0

with $k$ 1 for numerical stability. This yields a scalar boundary descriptor that differentiates smooth from irregular contours (Pan et al., 10 Jul 2025).

The empirical motivation is clinically specific. Across real BUS masks, benign tumors have $k$ 2, whereas malignant tumors have $k$ 3, with $k$ 4. SCMG therefore introduces a curvature loss

$k$ 5

which pushes benign masks toward smoother boundaries and malignant masks toward more irregular, lobulated, or spiculated boundaries (Pan et al., 10 Jul 2025). Semantic control is limited to tumor type $k$ 6 at the mask stage; other BI-RADS-style attributes are handled by the text-guided diffusion model to avoid conflicting shape guidance.

A common misconception is that “curvature-aware” must always mean an explicit $k$ 7-based term. The literature in fact shows multiple approximations. MaskFactory does not use an explicit curvature loss; instead, it enforces geometry through edge extraction, structural graphs, graph-space adversarial training, and a structure loss, while the image stage adds Canny conditioning for boundary precision (Qian et al., 2024). Range-Edit regularizes LiDAR object masks through convex hulls in range-image space, arguing that hulls mitigate label noise, regularize shape, and provide a general object shape representation (Uppur et al., 21 Nov 2025). Image blending with automatic mask generation relies on SAM-derived object masks followed by erosion and dilation, a morphological opening that removes protrusions and smooths boundaries without invoking curvature explicitly (Xue et al., 2023). DiffEdit derives masks from score-field differences under altered text conditions, so its notion of “shape” is localized semantic change rather than contour differential geometry (Couairon et al., 2022).

These contrasts are methodologically important. In SCMG, curvature is a supervised structural prior. In several adjacent systems, geometry is implicit, statistical, or procedural. The broader field therefore contains both strict and loose realizations of the semantic-curvature idea.

5. Evaluation, Downstream Utility, and Empirical Behavior

SCMG is evaluated both as a mask generator and as a contributor to synthetic-image utility. In the dedicated mask-generator ablation, removing curvature loss yields nearly indistinguishable curvature statistics for the two classes: benign $k$ 8 and malignant $k$ 9. Adding curvature loss restores separation: benign $s$ 0 and malignant $s$ 1. The paper interprets this as evidence that $s$ 2 functions as a structural prior, enforcing smoothness in benign tumor shapes and irregularity in malignant ones, thereby aligning masks with BI-RADS morphological descriptors (Pan et al., 10 Jul 2025).

The downstream image generator uses SCMG-produced masks for synthetic BUS generation, after which the synthetic data are mixed with real training data at ratios of $s$ 3, $s$ 4, $s$ 5, and $s$ 6. The reported downstream tasks are benign-versus-malignant classification with ResNet50, DenseNet121, EfficientNet-B0, and MobileNetV2, and lesion segmentation with UNet, UNet++, ResUNet, and Attention UNet. The paper states that synthetic augmentation improves AUC/F1 in classification and DSC in segmentation across BUSI, UDIAT, BrEaST, BUS-UCLM, and QAMEBI, and also improves generalization to the external STU dataset in segmentation (Pan et al., 10 Jul 2025). The same study also reports quantitative evaluation on six public BUS datasets and visual Turing tests by experienced sonographers confirming realism.

The paper is careful about attribution: it does not isolate the contribution of SCMG from the rest of the BUS synthesis framework in the downstream experiments. What is isolated is the mask-level curvature ablation. This makes SCMG’s direct contribution clearest at the level of morphological plausibility, with downstream benefit inferred through the full generative pipeline rather than separately measured.

The broader literature shows comparable importance of high-quality mask generation. MaskFactory reports gains over DatasetDM and Dataset Diffusion on DIS5K metrics, and its mask similarity analysis gives CLIP cosine similarity values of $s$ 7 for rigid masks, $s$ 8 for non-rigid masks, and $s$ 9 for the combined setting (Qian et al., 2024). XMask3D reports that on ScanNet B12/N7, novel-class mIoU rises from $\hat{M}$ 0 for PLA to $\hat{M}$ 1 for XMask3D, with mask-level alignment and 3D-conditioned mask generation identified as key factors (Wang et al., 2024). These results do not establish a single universal metric for Semantic-Curvature Mask Generators, but they do support the general proposition that mask quality can materially affect downstream fidelity, robustness, and open-vocabulary transfer.

Several neighboring systems clarify what SCMG is not. The face-part generator of "Automatic Generation of Semantic Parts for Face Image Synthesis" (Fontanini et al., 2023) is semantically structured and capable of local shape manipulation through class-wise latent codes and bi-directional LSTMs, but it explicitly lacks curvature losses and exhibits smoothing of high-frequency details such as hair contours. ASMA for adversarial face forgery attacks constructs semantic masks from face parsing and CAM-based region selection, but its geometry comes from semantic segmentation coherence and perturbation restriction rather than curvature-aware mask optimization (Liu et al., 2024). XMask3D introduces geometry-aware mask generation through 3D-conditioned diffusion and mask-level cross-modal reasoning, yet its curvature discussion remains prospective rather than implemented (Wang et al., 2024).

The limitations of explicit SCMG are correspondingly clear. In the BUS formulation, conditioning is restricted to tumor type and bounding rectangle; other shape descriptors such as margin subtype, multiplicity, or finer BI-RADS morphology are not directly encoded at the mask stage (Pan et al., 10 Jul 2025). The method is designed for 2D BUS rather than 3D ultrasound. Its target curvature values are extracted from existing annotations, so bias or inconsistency in real masks can propagate into the prior. Mean absolute curvature is the sole geometric descriptor, leaving open the possibility of richer priors such as multi-scale curvature histograms, skeleton features, or learned contour priors.

A broader source of confusion concerns the balance between semantic controllability and geometric realism. Increasing semantic specificity can fragment the conditioning space; increasing geometric regularization can oversmooth or constrain diversity. The literature suggests multiple responses rather than a settled solution: graph-based topology constraints in MaskFactory (Qian et al., 2024), convex-hull supports in range-image space for LiDAR editing (Uppur et al., 21 Nov 2025), latent prompt-difference masks in DiffEdit (Couairon et al., 2022), and cross-modal mask-level alignment in XMask3D (Wang et al., 2024). This suggests that the Semantic-Curvature Mask Generator is less a single architecture than a design pattern whose explicitness varies by domain.

In that pattern, the defining ambition remains stable: masks should not merely mark regions of interest, but should encode semantically meaningful and geometrically plausible structure that downstream generative or recognition models can trust. SCMG in breast ultrasound is the clearest explicit realization of that ambition currently represented in the cited literature, because it couples class-conditioned mask synthesis with a differentiable, statistically validated curvature prior and integrates the resulting masks into a clinically controllable diffusion pipeline (Pan et al., 10 Jul 2025).