Adaptive Shape-Aware Boundary Enhancement

• Adaptive Shape-Aware Boundary Enhancement is a technique that integrates explicit shape priors with boundary-sensitive operators to distinctly model complex, ambiguous edges in structured data.
• It employs adaptive convolutional modules and attention mechanisms to fuse local geometric adaptivity with anisotropic feature sensitivity, ensuring precise segmentation.
• Empirical results, including improved Dice scores and reduced prediction errors, confirm its efficacy in both medical image segmentation and CFD domain reconstruction.

Adaptive Shape-Aware Boundary Enhancement (ASBE) encompasses a class of algorithmic modules and strategies designed to improve the modeling, detection, and fidelity of boundaries in structured data, primarily focusing on medical image segmentation and computational fluid dynamics (CFD) domain reconstruction. The central objective is to combine local geometric adaptivity (shape-awareness) with fine-scale or anisotropic feature sensitivity (boundary-awareness). This approach is motivated by the observation that classical convolutional or mask-based routines often blur or misrepresent highly variable, low-contrast, or concave structures, particularly for domains where edge delineation and morphological integrity are critical.

1. Foundational Principles and Motivation

ASBE targets the prevalent limitations in standard convolutional segmentation or reconstruction pipelines, where isotropic receptive fields and generic loss formulations often fail to enforce the correct morphology in regions of subtle or ambiguous boundaries. In medical imaging, this manifests as missed or imprecise organ boundaries; in CFD, as the spurious activation of nonphysical voxels or loss of domain features. The key principle is to infuse explicit shape priors and dynamic, direction-sensitive mechanisms into the learning or reconstruction process, thus aligning low-level feature representations or grid masks with the true object or domain geometry before major architectural stages or downstream computations (Qu et al., 3 Nov 2025, Sharifi et al., 17 Feb 2026, Park et al., 2021).

2. ASBE Methodologies Across Domains

Medical Image Segmentation

• U-Net Stems with ASBE: In the RDTE-UNet framework, ASBE replaces the standard convolutional stem. The module combines a channel-compression stage, a shape-prior pooling operation, and an adaptive rectangular convolution ("ARConv") with per-location, anisotropic sampling offsets derived from soft pooled feature statistics. Boundary accentuation is implemented via a simple local minus pooled-feature difference, optionally replaced by a morphological gradient. The outputs (shape-aware features and detail accentuation maps) are fused by concatenation and 1×1 convolution for channel mixing. This injects both morphological adaptivity and enhanced edge evidence into feature hierarchies before downsampling (Qu et al., 3 Nov 2025).
• Shape-Aware Attention in Cardiac Segmentation: The Contour- and Distance-Transform-Guided Shape-Aware Attention ("CDA-Net") approach employs two auxiliary V-transition streams: a Contour Transition Network (CTN) and a Distance Transform Transition Network (DTTN). The CTN extracts per-voxel contour probabilities from shallow features, while the DTTN regresses per-voxel distance transforms from deep features. These are combined in a shape-aware attention block that modulates decoder features by concatenating decoder, contour, and distance-transform tensors and passing them through convolution+squashing, yielding an attention map that enhances edge and shape fidelity. A penalty energy ensures contour responses align with interior boundaries as expressed by DT supervision (Park et al., 2021).

CFD Domain Reconstruction

• Topological Mask Recovery: ASBE addresses the challenge of reconstructing exact CFD flow domains from scattered point sets interpolated onto grids, which otherwise create convex hull artifacts and nonphysical regions. The strategy involves a fast distance-based masking (computing and thresholding the nearest-neighbor distance field), an adaptive α-shape method (filtering Delaunay simplices using a resolution-normalized radius parameter to recover true geometry), and a lightweight boundary inflation step (mask dilation at sub-voxel scales) to correct sampling gaps (Sharifi et al., 17 Feb 2026).
• Boundary Correction in Neural Operators: In neural operator architectures for PDE-based simulation, ASBE modules (e.g., GeoABC) inject boundary-aware corrections into intermediate latent states. This feature utilizes a geometry cache (signed-distance, normal, tangent, curvature) to enable a local tangent–normal decomposition of latent anchors, directional corrections modulated by geometry-aware gates, and a spatial confidence window that restricts correction to near-wall regions. The correction terms are weighted, projected back into latent space, and combined residually, significantly reducing near-boundary prediction error (Zhang et al., 8 Jun 2026).

3. Mathematical and Algorithmic Foundations

Across implementations, ASBE incorporates:

• Shape-Prior Pooling & Adaptive Convolution: Computation of low-rank statistics (e.g., average pooling) to capture coarse morphology, with subsequent use in predicting per-location sampling offsets for rectangular or anisotropic convolutions (Qu et al., 3 Nov 2025).
• Edge-Accentuating Operators: Differences between feature maps and their pooled versions (or morphological gradient operations) serve as a lightweight, learnable mechanism for edge highlighting.
• Attention and Modulation Blocks: Fusion of shape/edge cues with deeper features through attention mechanisms—either as channelwise weights derived from auxiliary predictions (contour, distance transform), or as spatial gates derived from distance fields or confidence maps (Park et al., 2021, Zhang et al., 8 Jun 2026).
• Mask Generation Based on Distance and α-Shapes: Construction of binary masks using nearest-neighbor distance fields and adaptive α-shape reconstruction with triangulation and circumradius filtering based on local resolution. Post-processing with dilation recovers samples missed due to discretization or under-sampling (Sharifi et al., 17 Feb 2026).
• Boundary Correction in Neural Operators: Latent representations are projected into boundary-aligned (tangent/normal) frames, and pointwise corrections are synthesized by MLPs, gated by geometry-derived weights. These are recombined into global latent space and integrated into the backbone features as a residual update (Zhang et al., 8 Jun 2026).

4. Implementation Details and Hyperparameters

RDTE-UNet/ASBE Module (Qu et al., 3 Nov 2025):

• Layer sequence: 1×1 conv + BN + ReLU → avg pool → OffsetMLP → ARConv → difference map → 3×3 conv + BN + ReLU → concat → 1×1 conv + BN + ReLU.
• Kernel/offset settings: ARConv uses (3×5) and (5×3) kernels; learned offsets via two-layer MLP with SiLU activations.
• Hyperparameters: Channel compression to C_in/2, ARConv outputs C₂=C₁, final output C_in. Weight decay 1e–4, SGD momentum 0.9 or Adam (lr=1e–4).
• Additional supervision: BCE loss on predicted vs. ground-truth edges (GT_edge from morphological gradient), with λ_bd=0.1.

Cardiac Segmentation Shape-Aware Attention (Park et al., 2021):

• Supervision terms: Dice loss (segmentation), boundary BCE loss, distance-transform MSE, penalty energy; λ_1=1, λ_2=20, λ_3=10, λ_4=1.
• Auxiliary modules: Shallow encoder to CTN for boundary, deep encoder to DTTN for shape; both lightweight, class-specific branches.

CFD Reconstruction (Sharifi et al., 17 Feb 2026):

• Distance-based masking: Distance computation vs. scattered P; τ (threshold) is the minimum grid spacing; morphological closing with structure element radius r_c=5.
• Adaptive α-shape: α = β·ē, β=1 suffices multiple geometries; dual criteria for simplex retention and mask extraction.
• Boundary inflation: Dilation with η=1.002, affecting retention by up to ≈3%, with ghost activation increase <0.08%.

Neural Operator Boundary Correction (Zhang et al., 8 Jun 2026):

• Geometry cache: For each location, precompute the signed-distance, normal, tangent, curvature, window w(x), and confidence c(x).
• Injection layer: Inserted at backbone mid-depth (layer l*), adds ≤2% parameter overhead.
• Directional gates: α_t, α_n, α_m from 1-layer MLP.
• MLP details: Correction branches are 2-layer (width=32, ReLU), context dim d_h≈32–64.
• Losses: Combine global field L2, boundary-weighted L2, anchor-alignment L2, with application-specific weights.

5. Empirical Results and Impact

• Segmentation Accuracy: In RDTE-UNet, ASBE improves Synapse dataset performance from DSC=84.97% and HD95=15.17 mm to DSC=86.63%, HD95=11.69 mm (ASBE adds +1.66% Dice, –3.48 mm HD95 vs. baseline). Qualitative differences include continuous detection of thin vessels and boundary lobulations, with substantially fewer spurious holes (Qu et al., 3 Nov 2025).
• Cardiac Segmentation: Shape-aware contour attention yields a Dice coefficient of 0.8700 (vs. 0.8184–0.8203 for attention and shape-aware U-Nets), with improvements in ASSD, precision, sensitivity, and reduction of false positives at organ interfaces (Park et al., 2021).
• CFD Domain Recovery: Distance-based ASBE masking generates masks in 15–18 ms (500–800× faster than classical α-shape); adaptive α-shape is 1.7–2.6× faster than classical. Retention after minimal dilation is up to 2.96% improved, with ghost fraction <0.08% (see Table below) (Sharifi et al., 17 Feb 2026).

Method	Speedup (vs. α-shape)	Retention Increase	Ghost Increase
Distance-based ASBE	500–800×	up to 2.96%	≤0.06%
Adaptive α-shape (β=1)	1.7–2.6×	up to 2.59%	≤0.08%

• Neural Operators: On 2D airfoil, GeoABC reduces near-wall relative L2 error by ≈72% (0.1320→0.0369) and pressure MAE by ≈80% (0.2085→0.0426) (Zhang et al., 8 Jun 2026).

6. Extensions, Limitations, and Domain-Specific Variations

ASBE modules have been widely adapted to the context of the task. In segmentation, modules are often placed at the entry point of the network or as intermediate attention gates. In CFD and neural-PDE settings, ASBE serves as a mask generator and boundary modulator, with correction localized to near-boundary regions using geometric priors.

While ASBE dramatically improves boundary localization and shape integrity, explicit shape and edge supervision typically requires precomputing ground-truth contours/distance fields, and the adaptability of anisotropic mechanisms is sensitive to the specificity and variability of training domains. For highly fragmented or multi-organ targets, extension of shape/prior modules must address increased complexity in geometric topology and class-wise partitioning. The adaptive tuning of correction or mask parameters alleviates dataset-specific manual tuning, enabling broader applicability with minimal re-engineering.

7. Significance and Outlook

ASBE unifies morphological adaptivity and boundary sharpness, substantiated by statistically significant empirical improvements across both medical imaging and physical domain modeling. Its modular design—leveraging adaptive convolutions, shape-aware attention, geometric masking, and tangent–normal decomposition—facilitates integration into diverse deep learning and neural operator pipelines. ASBE's demonstrated reproducibility, low overhead, and strong boundary fidelity make it a foundational building block for segmentation and reconstruction pipelines where geometric precision and robust boundary delineation are critical (Qu et al., 3 Nov 2025, Park et al., 2021, Sharifi et al., 17 Feb 2026, Zhang et al., 8 Jun 2026).