Semantic-Aware Deformable Feature Alignment for Thyroid Nodule Segmentation in Ultrasound

• CNN denotes the convolutional neural network used for high-resolution local feature extraction, while ViT models relations between image patches.Source: Glossary · p. 2
• CLCA names SDFA-Net’s Clustered Linear Context Aggregation module, and DCN refers to a deformable convolutional network that learns spatial sampling offsets.Source: Glossary · p. 2
• BCE is a pixel-level segmentation loss; DSC measures region overlap, and HD95 measures boundary distance while reducing extreme-outlier influence.Source: Glossary · p. 2
• GFLOPs reports computational cost, while FPS reports inference speed.Source: Glossary · p. 2

• Low contrast, speckle, acoustic shadowing, feature overlap, and geometric variability make thyroid nodule boundaries difficult to delineate.Source: §1 · p. 2
• Fixed spatial fusion assumes global and local features correspond, an assumption that can fail when irregular nodules geometrically misalign them.Source: §1 · p. 3
• SDFA-Net uses clustered linear context aggregation and dynamic prototypes to filter global features, then deformably aligns the selected context with local boundaries.Source: §1 · p. 3
• SDFA-Net records the best Dice and HD95 results on TN3K, TUCC, and DDTI, and retains strong performance on irregular and small nodules.Source: §1 · p. 3
• Experiments compare ten segmentation models using Dice and HD95, alongside component studies and shape-and-size analyses.Source: §1 · p. 3

2.1. Context-Aware Segmentation Architectures

• CNNs capture local texture but have limited receptive fields, while Transformers model global context yet may mix noisy background tokens and incur quadratic attention costs.Source: §2.1 · pp. 3–4

2.2. Semantic Filtering and Deformable Feature Fusion

• Attention gates weight feature regions by local appearance, whereas prototype learning provides class-specific representations from global feature statistics.Source: §2.2 · p. 4
• Deformable convolution learns input-specific sampling offsets, adapting feature integration to irregular geometry where fixed-grid fusion can fail.Source: §2.2 · p. 4
• Figure 1 depicts a dual-path design that filters semantically irrelevant features before deformably aligning global context with local boundaries.Source: Figure 1 · p. 5

3.1. Overview of SDFA-Net Architecture

• A dual-path architecture preserves high-resolution local details while refining global context with dynamic prototypes before deformable alignment to local geometry.The local path uses a ResNet-34 backbone; the global path captures long-range anatomical dependencies.Source: §3.1 · p. 5

3.2. Dual-Path Feature Extraction and Context Aggregation

• A ResNet-34 dual path preserves multi-scale local features while CLCA groups similar global tokens to model long-range dependencies at O(NKC).CLCA uses Sinkhorn-normalized affinities to balance cluster assignments.Source: §3.2 · pp. 6–7

3.3. Affinity-Driven Token Modulation

• During training, masked averages of normalized global tokens form class prototypes that gate tokens by foreground affinity; inference uses running-average prototypes.The prototype memory is updated with an exponential moving average because ground-truth masks are unavailable at inference.Source: §3.3 · pp. 8–9

3.4. Spatially-Aligned Feature Fusion

• At strides 16, 8, and 4, predicted dense offsets deformably warp semantic features toward local geometry, with stride-4 fusion feeding the segmentation head.The cascaded alignment targets irregular boundaries where rigid transformations may fail.Source: §3.4 · p. 10

3.5. Loss Function and Training Objective

• Training combines BCE/Dice segmentation loss for class imbalance with prototype contrastive regularization that separates foreground and background tokens (λ=0.1).The prototype loss pulls foreground tokens toward the foreground prototype and pushes background tokens away from the background prototype.Source: §3.5 · pp. 11–12

4.1. Datasets and Implementation Details

• The study evaluates TN3K, DDTI, and TUCC using 70/10/20 training, validation, and test splits, with all images resized to 512 × 512 and normalized.TN3K contains approximately 3,500 static images from 208 patients; DDTI derives from 61 patient video sequences, while TUCC provides cine-clips acquired with different equipment and protocols.Source: §4.1 · p. 12
• Evaluation uses Dice Similarity Coefficient (DSC) for region overlap and HD95 for boundary fidelity, with lower HD95 indicating better boundary alignment.HD95 removes the top 5% of surface-distance outliers to reduce sensitivity to noise.Source: §4.1 · p. 12
• Training uses a ResNet-34 pretrained on ImageNet for local features, AdamW optimization, and a polynomial learning-rate schedule over 120 epochs.The global branch starts with random weights; training uses a composite BCE-and-Dice segmentation loss with prototype contrastive loss and augmentation including rotations, flips, elastic deformations, and Gaussian noise.Source: §4.1 · pp. 12–13

4.2. Ablation Studies

• The ablations on TN3K examine each module’s contribution, design alternatives, and sensitivity to key hyperparameters.The study uses the primary dataset for these component analyses.Source: §4.2 · p. 13

4.2.1. Component Ablations

• 87.42±0.96% Dice and 12.37 pixels HD95 are the full SDFA-Net’s best component-ablation results on TN3K.Adding context aggregation, token modulation, and spatially aligned fusion progressively improves the U-Net baseline.Source: Table 1 · p. 13, §4.2.1 · p. 13

4.2.2. Global Context Aggregation Method

• At K = 8, clustered context aggregation achieves 83.46% Dice, 16.83 px HD95, and 35.11 GFLOPs, balancing performance and efficiency.Self-attention reaches 83.51% Dice but costs 48.73 GFLOPs, approximately 39% more; performance saturates at K = 16.Source: Table 2 · p. 14, §4.2.2 · p. 14

4.2.3. Prototype Modulation Strategy

• Dynamic prototypes outperform static prototypes by 0.81% Dice on TN3K and 2.04% Dice on TUCC.Removing modulation reduces TN3K Dice by 2.27%; the authors attribute the cross-dataset gains to input-specific semantic filtering.Source: Table 3 · p. 14, §4.2.3 · p. 14

4.2.4. Feature Fusion Strategy

• Deformable alignment improves Dice by an additional 1.14% over rigid alignment and achieves the lowest HD95, 12.37 pixels.Rigid alignment improves Dice by 0.55% and HD95 by 0.57 pixels over concatenation; the authors cite non-rigid nodule shapes as the reason dense offsets matter.Source: Table 4 · p. 15, §4.2.4 · p. 15

4.2.5. Hyperparameter Sensitivity

• At λ = 0.10, the prototype loss yields 87.42% Dice and 12.37 pixels HD95; removing it decreases Dice by 0.55% and increases HD95 by 0.61 pixels.Larger weights bring diminishing returns and slight degradation.Source: Table 5 · p. 15, §4.2.5 · p. 15

4.3. State-of-the-Art Comparisons

• All three datasets use supervised training with a consistent 70/10/20 split, and performance is measured by Dice and HD95.The datasets are evaluated independently; TUCC tests data acquired with different equipment and imaging protocols.Source: §4.3 · pp. 15–16

4.3.1. Performance on TN3K

• TN3K: SDFA-Net scores 87.42% Dice, 1.41% above AVFM, and reduces HD95 to 12.37 pixels from AVFM’s 15.42 pixels.The qualitative results also report an advantage on irregular nodules through prototype modulation that suppresses irrelevant features.Source: §4.3.1 · p. 16, Table 6 · p. 16, Figure 5 · p. 17

4.3.2. Performance on TUCC

• TUCC: SDFA-Net achieves 71.83% Dice, surpassing the best cross-dataset baseline, AVFM, by 2.96%.The authors report that affinity-driven token modulation suppresses protocol-specific artifacts and background textures.Source: §4.3.2 · pp. 16–17, Table 7 · p. 17

4.3.3. Performance on DDTI

• DDTI: SDFA-Net achieves 86.73% Dice, 1.53% above HFA-UNet, and an HD95 of 11.05 pixels.The reported improvement is most pronounced on nodules with low annotation quality, where baseline Dice varies more.Source: §4.3.3 · pp. 17–18, Table 8 · p. 18

Analysis setup

• Solidity is defined as nodule area divided by convex-hull area; values nearer 1 indicate regular, convex shapes, while lower values indicate irregular boundaries.The paper associates such lower-solidity shapes with lobulated or spiculated boundaries.Source: §4.4 · p. 18
• Test samples are divided into three equal-sized solidity groups and three pixel-area groups using terciles, and Figure 6 presents subgroup Dice distributions.The solidity groups are Highly Irregular, Moderately Irregular, and Regular; the size groups are Small, Medium, and Large.Source: §4.4 · p. 19

Results

• SDFA-Net outperforms all baselines across solidity and size groups, with gains emphasized for highly irregular and small nodules.The authors report consistently reduced HD95 across all three datasets.Source: §4.4 · p. 19

Design interpretation

• Filtering semantic noise before fusion and aligning the remaining context with local boundary evidence links the two stages to gains in region overlap and boundary distance.The reported pattern is strong for irregular and small nodules.Source: §5 · p. 19
• The model first selects global tokens belonging to the target class, then determines where to place that information in local features, keeping noisy context from steering offsets.The discussion suggests this order may apply to dense recognition when class evidence is stable at low resolution but geometry varies at high resolution.Source: §5 · p. 19

Clinical perspective

• At approximately 23 frames per second, SDFA-Net is described as a candidate for computer-aided diagnosis, with robust handling of speckle and shadowing potentially reducing manual correction.Consistent performance across equipment and protocols is presented as supporting utility across clinical centers.Source: §5 · p. 19

Scope and future work

• The current evidence covers fully supervised 2D ultrasound segmentation on three public datasets; future work will examine sparse annotation, volumetric ultrasound, and deployment-specific optimization.These settings would test the approach under weaker supervision, richer spatial context, and tighter computation budgets.Source: §5 · p. 19

6. Conclusion

• SDFA-Net builds global context, filters context that does not match the nodule class using dynamic prototypes, and registers the result with local features across three resolutions.The conclusion frames this sequence as providing semantic context that agrees with target geometry.Source: §6 · p. 20
• On TN3K, TUCC, and DDTI, SDFA-Net achieves the highest Dice and lowest HD95 among compared methods, with strong benefits reported for highly irregular and small nodules.The conclusion also reports component studies tracing gains to the context, prototype, and alignment modules, and inference at about 23 frames per second.Source: §6 · p. 20
• The study covers fully supervised 2D ultrasound segmentation; sparse labels, 3D ultrasound, and deployment-specific optimization remain future tests.The conclusion presents semantic relevance and spatial correspondence as linked decisions for dense prediction across scales.Source: §6 · p. 20

• Lei Zhang handled conceptualization, methodology, software, and original-draft writing; Ruhuan Zhou handled software, validation, data curation, and original-draft writing.Source: CRediT authorship contribution statement · p. 20
• Busheng Tong contributed methodology and data curation; Dan Lv and Bingyu Liang contributed visualization.Source: CRediT authorship contribution statement · p. 20
• Taibao Wang contributed original-draft writing; Yifan Gao contributed software and visualization; Xinyu Shi contributed supervision.Source: CRediT authorship contribution statement · p. 20
• Ming Zhang contributed review and editing and supervision; Qin Wang contributed review and editing, supervision, and funding acquisition.Source: CRediT authorship contribution statement · p. 20