Organizations: Department of Computer Science and Engineering, United International University, United City, Madani Ave, Dhaka 1212, Bangladesh · Faculty of Informatics, Institute of Industry-Academia Innovation, Department of Data Science and Engineering, Eötvös Lor´and University, P´azm´any P´eter S´et´any 1/C, Budapest 1117, Hungary · Department of Computer Science and Engineering, BRAC University, Pragati Sarani, Merul Badda, Dhaka 1212, Bangladesh
Accurate thyroid ultrasound segmentation is often challenged by low contrast, speckle noise, and unclear boundaries. Although recent methods have improved segmentation accuracy, many rely on resource-intensive architectures or lack explicit integration of multiscale features with global biomedical visual guidance. In this paper, we introduce ThyCLIPNet, a lightweight semantic-guided hybrid encoder-decoder framework that integrates BiomedCLIP-derived biomedical semantic guidance into a lightweight multi-scale CNN segmentation pipeline. The encoder integrates MobileNetV2 with efficient channel attention, while atrous spatial pyramid pooling and a custom convolutional block attention module enrich bottleneck features. The decoder combines hierarchical skip connections and lightweight attention refinement with a BiomedCLIP-guided gated fusion pathway that projects vision-only global biomedical embeddings into decoder feature space and selectively integrates them through semantic-local fusion and spatial gating. To the best of our knowledge, ThyCLIPNet is among the first lightweight thyroid ultrasound segmentation frameworks to use BiomedCLIP's vision encoder alone for image-only global semantic guidance without text prompting. Experiments on TG3K, TN3K, DDTI, and PKTN achieve dice similarity coefficients of 96.22%, 87.58%, 84.73%, and 80.70%; intersection over union scores of 92.72%, 77.91%, 73.51%, and 67.64%; and 95th-percentile hausdorff distances of 3.75, 16.38, 18.23, and 10.86, respectively. ThyCLIPNet uses 8.55M parameters and 22.99G FLOPs. Overall, the results support integrating global biomedical semantic guidance with lightweight multi-scale CNN representations for robust and computationally efficient thyroid ultrasound segmentation. Source code: https://github.com/Tasnim-Jahan/ThyCLIPNet. [Abstract shortened for arXiv. See PDF for full abstract.]
Figures & tables
Dataset
Total Images (Original)
Train Images (Original)
Validation Images
Test Images
Augmentation Factor (k)
Augmented Train Images
Final Train Images
Total Images (After Augmentation)
Ultrasound Devices
TG3K [ 62 ]
3,585
2,868
358
359
1
2,868
5,736
6,453
GE Logiq E9
TN3K [ 63 ]
3,493
2,304
575
614
1
2,304
4,608
5,797
GE Logiq E9; ARIETTA 850; RESONA 70B
DDTI [ 64 ]
637
509
63
65
5
2,545
3,054
3,182
Toshiba Nemio 30; Toshiba Nemio MX
PKTN [ 54 ]
1,005
804
100
101
4
3,216
4,020
4,221
Not disclosed
Table 1: Key characteristics and dataset splits for model evaluation.
Figure 1: Illustration of the preprocessing and data‐augmentation pipeline: ultrasound images and masks are resized to 256 × 256 and normalized; dataset-specific offline augmentations are applied; 512-dimensional BiomedCLIP embeddings are extracted from original and augmented images; dynamic online augmentations are performed on both images and embeddings; and the resulting data are fed into the segmentation model for training.
Figure 2: Architecture of the proposed ThyCLIPNet framework for thyroid ultrasound segmentation, illustrating: (a) MobileNetV2 encoder integrated with ECA modules; (b) multi-scale context capture using an ASPP block fused with a custom CBAM in the bottleneck; (c) global semantic guidance from BiomedCLIP image embeddings integrated through the BiomedCLIP-Guided Gated Fusion (BGF); and (d) three-stage U-Net–style decoder with skip connections and lightweight CBAM for progressive boundary refinement.
Figure 3: Feature bottleneck module: the ECA-refined encoder output F4′ is processed by ASPP block with SpatialDropout to yield FASPP , which is then refined by custom CBAM to produce FCustomCBAM before decoding.
Figure 4: Lightweight CBAM attention block: decoder feature maps first pass through channel attention (global average pooling → dense → relu → dense → sigmoid) and then spatial attention (pooled features concatenated →7×7 convolution → sigmoid) to selectively refine and enhance informative regions before further decoding.
Figure 5: BiomedCLIP-guided Gated Fusion (BGF) module: global biomedical image embeddings are projected, reshaped, and bilinearly upsampled, then combined with decoder features using weighted addition and refined with spatial attention gating before concatenation.
Dataset
Methods
DICE (%) ↑
IoU (%) ↑
Accuracy (%) ↑
Recall (%) ↑
Precision (%) ↑
HD95 ↓
TG3K [ 62 ]
Coarse for Fine [ 73 ]
80.67 ± 1.09
—
—
—
—
16.63 ± 0.84
Weakly‐supervised Framework [ 74 ]
85.92 ± 1.53
—
93.00
86.00
91.00
13.09 ± 0.43
Adapting Vision [ 75 ]
83.04
—
—
—
—
12.55
SAMTNU [ 83 ]
83.87
73.79
97.64
85.17
—
23.98
DCDT [ 92 ]
82.04 ± 1.11
69.55 ± 1.85
—
—
—
26.68 ± 0.74
FMF-Net [ 91 ]
92.20
—
—
88.90
—
0.94
Table 2: Table 2 . Comparison of segmentation effects between the proposed and SOTA methods across all datasets.
Figure 6: Multi‐metric bubble‐plot comparison of ThyCLIPNet versus state‐of‐the‐art methods on three thyroid ultrasound datasets (TG3K: top row; TN3K: middle row; DDTI: bottom row). The left column shows DICE versus HD95 (bubble size ∝1/HD95 ); the center column shows DICE versus Accuracy (bubble size ∝Accuracy ); and the right column shows dataset‐specific comparisons: Accuracy versus HD95 for TG3K (bubble size ∝1/HD95 ), DICE versus IoU for TN3K and DDTI (bubble size ∝IoU ). Bubble color represents the DICE score.
Figure 7: Qualitative visualization of absolute segmentation error maps ( ∣GT−Prediction∣ ) on TG3K, TN3K, DDTI, and PKTN for ThyCLIPNet and eight ablated variants. Each row shows absolute error, highlighting boundary and shape discrepancies.
Figure 8: Qualitative prediction matrix on representative test images from TG3K, TN3K, DDTI, and PKTN. Rows correspond to datasets; columns show the input ultrasound image, ground truth (GT), and predicted masks from ThyCLIPNet, BPAT-UNet, TRFE++, CLIP-TNseg, and UNet++.
Figure 9: Representative hard-case segmentation results across four thyroid ultrasound datasets (TG3K, TN3K, DDTI, PKTN). Each row depicts a difficult example characterized by low contrast, boundary ambiguity, or structural noise. Columns show (from left to right): input image with annotated contours (ground truth in orange, prediction in blue), ground truth mask, predicted mask, and absolute error overlay. DICE and HD95 are reported per case to illustrate how boundary ambiguity elevates HD95 and how errors are typically concentrated near challenging edges.
Dataset
Architecture
DICE (%) ↑
IoU (%) ↑
Accuracy (%) ↑
Recall (%) ↑
Precision (%) ↑
HD95 (px) ↓
TG3K [ 62 ]
ThyCLIPNet w/o ECA
96.08 ± 0.10
92.45 ± 0.20
99.31 ± 0.00
97.07 ± 0.20
95.11 ± 0.30
3.63 ± 0.46
ThyCLIPNet w/o ASPP
96.12 ± 0.20
92.53 ± 0.40
99.32 ± 0.00
96.87 ± 0.50
95.38 ± 0.50
3.96 ± 0.23
ThyCLIPNet w/o BiomedCLIP
94.21 ± 0.20
89.06 ± 0.40
98.94 ± 0.00
98.34 ± 0.40
90.43 ± 0.60
4.94 ± 0.42
ThyCLIPNet w/o custom CBAM
95.91 ± 0.60
92.15 ± 1.10
99.28 ± 0.10
96.80 ± 0.70
95.04 ± 0.80
4.02 ± 0.19
ThyCLIPNet w/o lightweight CBAM
96.36 ± 0.10
92.97 ± 0.20
99.36 ± 0.00
96.95 ± 0.10
95.78 ± 0.20
3.79 ± 0.23
ThyCLIPNet w/o both CBAM variants
96.54 ± 0.20
93.32 ± 0.30
99.39 ± 0.00
97.18 ± 0.10
95.92 ± 0.20
3.75 ± 0.13
Table 3: Table 3 . Ablation‐study results on four datasets (TG3K, TN3K, DDTI, and PKTN). Metrics are mean ± standard deviation over five independent runs. DICE, IoU, accuracy, recall, precision and HD95 are shown.
Variants
Params
Δ Params (%)
Train Time (s)
Δ Train (%)
Inf Time (s)
Δ Inf (%)
w/o ECA
8,550,704
0.00
16,070.82
+5.93
0.01305
+2.84
w/o BiomedCLIP
7,821,736
-8.53
8,479.56
-44.11
0.01148
-9.54
w/o lightweight CBAM
8,312,114
-2.79
15,083.26
-0.58
0.01150
-9.38
w/o both CBAM variants
8,270,353
-3.28
20,043.51
+32.12
0.01267
-0.16
w/o (BiomedCLIP+ECA)
7,821,720
-8.53
18,130.88
+19.51
0.01874
+47.68
w/o (BiomedCLIP+ASPP)
4,543,656
-46.86
14,859.38
-2.05
0.00879
-30.73
Table 4: Table 4 . Efficiency comparison on TG3K. Values are reported as the mean over five independent runs. Δ values are relative to the proposed ThyCLIPNet model.
Dataset
Architecture
DICE (%)
IoU (%)
Accuracy (%)
Recall (%)
Precision (%)
HD95 (px)
TG3K
ThyCLIPNet+TTA
96.43 ± 0.10
93.10 ± 0.10
99.37 ± 0.00
97.03 ± 0.20
95.83 ± 0.20
3.69 ± 0.16
TN3K
ThyCLIPNet+TTA
87.92 ± 0.20
78.45 ± 0.30
97.03 ± 0.10
89.49 ± 0.70
86.42 ± 0.80
16.09 ± 0.88
DDTI
ThyCLIPNet+TTA
85.05 ± 0.50
73.99 ± 0.80
95.35 ± 0.20
89.72 ± 1.40
80.88 ± 1.80
18.41 ± 1.28
PKTN
ThyCLIPNet+TTA
80.77 ± 0.30
67.75 ± 0.40
98.68 ± 0.00
80.45 ± 1.80
81.17 ± 2.00
10.91 ± 0.51
Table 5: Table 5 . Performance of ThyCLIPNet + TTA across four datasets metrics as mean ± std.
Model
Params (M)
Inference Time (s/img)
FPS
FLOPs (G)
GPU Mem (GB)
UNet++ [ 95 ]
9.16
0.01359
73.58
69.43
1.81
TRFE+ [ 77 ]
45.44
0.03834
26.34
160.77
0.76
BPAT-UNet [ 81 ]
72.39
0.07648
13.08
128.66
0.55
DeepLabV3+ with MobileNetV2 [ 65 ]
5.85
0.01223
81.77
12.37
0.77
U-Net with MobileNetV2 [ 28 ]
3.75
0.00698
143.27
15.85
0.52
ThyCLIPNet-Proposed
8.55
0.01269
78.81
22.99
0.76
Table 6: Table 6 . Computational benchmark comparison of ThyCLIPNet with representative baselines. Representative lightweight models were re-implemented under identical evaluation settings (input size = 256 × 256, batch size = 1), while remaining architectures were benchmarked under the same hardware and inference environment. FLOPs and parameter counts are architecture-dependent and dataset-independent. Inference times represent median (P50) values over five independent runs; for ThyCLIPNet, computational measurements correspond to the segmentation stage using precomputed BiomedCLIP embeddings. FPS is computed as 1/time .
Model
Params (M)
Inference Time (s/img)
FPS
FLOPs (G)
GPU Mem (GB)
GDSSA-Net [ 53 ]
35.00
0.042
—
—
—
MSAC-DD [ 52 ]
25.70
0.233
—
—
2.55
FCG-Net [ 90 ]
11.35
—
20.25
—
—
FDE-Net [ 97 ]
—
0.036
—
—
—
DualSwinUNet++ [ 98 ]
—
0.180
—
—
—
TNSNet [ 96 ]
46.47
0.080
12.5
—
25.42
Table 7: Table 7 . Comparison with previously reported thyroid segmentation networks. Values are primarily extracted from the corresponding original publications, whereas ThyCLIPNet was benchmarked under the experimental protocol adopted in this study (input size = 256 × 256, batch size = 1).
Figure 10: Six‐panel Grad-CAM interpretability visualizations for two representative cases from each thyroid ultrasound dataset (TG3K: top two rows; TN3K: next two rows; DDTI: next two rows; PKTN: bottom two rows). From left to right: (i) input image, (ii) ground-truth mask, (iii) predicted mask, (iv) Grad-CAM heatmap, (v) overlay of prediction, and (vi) overlay of Grad-CAM heatmap.
Thyroid nodules affect a large proportion of the general population. Although most are benign, a non-negligible fraction is malignant. Ultrasound imaging is the primary diagnostic modality, and accurate nodule segmentation is a prerequisite for CAD systems to classify malignancy and assess metastatic risk.\ In this work, we propose a thyroid nodule segmentation framework that integrates a DeepLabV3+, channel-attention modules and post-hoc conformal prediction, providing both accurate mask and pixel-level uncertainty estimates. We address thyroid ultrasound nodule segmentation as the first stage of a pipeline that also includes malignancy classification and metastatic risk prediction. We adopt DeepLabV3+, whose Atrous Spatial Pyramid Pooling (ASPP) captures multi-scale context, with an EfficientNet-B5 backbone for its accuracy-efficiency trade-off. Additionally, we test three channel-attention modules: Squeeze-and-Excitation, CBAM and Grouped Channel Attention (GCA), paired with two loss formulations (Dice-BCE and Boundary).\ Since binary masks do not indicate prediction confidence, we wrap the trained model with a split conformal prediction layer that produces pixel-level uncertainty maps with statistical coverage guarantees.\ We evaluate all configurations on the TN3K public benchmark using a fixed split and directly compare them with state-of-the-art methods. The plain backbone already improves upon previously reported results, with CBAM + Dice-BCE emerging as the best overall configuration.
Francesco Teodori, Pierpaolo Brutti, Umberto Ferraro Petrillo
Department of Statistical Sciences, Sapienza University of Rome, Rome, Italy.
Deep learning (DL) methods dominate thyroid nodule segmentation in ultrasound (US) images, achieving high Dice scores but at the cost of millions of parameters, GPU-dependent training via backpropagation, and limited mathematical tractability. These limitations impede deployment in resource-constrained environments. In this paper, we propose MedSaab-US, a backpropagation-free segmentation framework grounded in the Green Learning paradigm. MedSaab-US extracts multi-scale spatial-frequency features by combining multi-level Discrete Wavelet Transform (DWT) with multi-scale channel-wise Saab (Subspace Approximation with Adjusted Bias) transforms at patch sizes of 5 x 5, 11 x 11, and 21 x 21 pixels. Label-Assisted Greedy (LAG) feature selection retains the most discriminative features, which are fed to an XGBoost classifier for pixel-wise prediction. The Saab transform parameters are determined analytically from data statistics, while XGBoost employs iterative greedy tree construction without requiring backpropagation. Evaluated on the TN3K dataset (2,879 training and 614 test images), MedSaab-US achieves a mean Dice coefficient of 0.4784 +/- 0.2190, precision of 0.5768, and recall of 0.5604, with a model footprint under 500K parameters and CPU-only inference in approximately 0.3 seconds per image. We present this result as an exploratory non-DL baseline for thyroid ultrasound segmentation and analyze the specific challenges posed by isoechoic nodules. An ablation study further quantifies the contribution of each pipeline component, including separate evaluations of LAG feature selection and training-set size.
Mohammad Amanour Rahman
Department of Computer Science and Engineering Ahsanullah University of Science and Technology (AUST) Dhaka, Bangladesh
Thyroid nodule classification using ultrasound imaging is essential for early diagnosis and clinical decision-making; however, despite promising performance on in-distribution data, existing deep learning methods often exhibit limited robustness and generalisation when deployed across different ultrasound devices or clinical environments. This limitation is mainly attributed to the pronounced heterogeneity of thyroid ultrasound images, which can lead models to capture spurious correlations rather than reliable diagnostic cues. To address this challenge, we propose PEMV-thyroid, a Prototype-Enhanced Multi-View learning framework that accounts for data heterogeneity by learning complementary representations from multiple feature perspectives and refining decision boundaries through a prototype-based correction mechanism with mixed prototype information. By integrating multi-view representations with prototype-level guidance, the proposed approach enables more stable representation learning under heterogeneous imaging conditions. Extensive experiments on multiple thyroid ultrasound datasets demonstrate that PEMV-thyroid consistently outperforms state-of-the-art methods, particularly in cross-device and cross-domain evaluation scenarios, leading to improved diagnostic accuracy and generalisation performance in real-world clinical settings. The source code is available at https://github.com/chenyangmeii/Prototype-Enhanced-Multi-View-Learning.
Yangmei Chen, Zhongyuan Zhang, Xikun Zhang +4
Guangdong Institute of Smart Education, Jinan University, Guangzhou 510632, China