Source-linked AI summary
AAU-net: An Adaptive Attention U-net for Breast Lesions Segmentation in Ultrasound Images
Gongping Chen, Yu Dai, Jianxun Zhang, Moi Hoon Yap
TL;DR
Breast-lesion segmentation in ultrasound is difficult because complex patterns, similar intensities, variable morphologies, and blurred boundaries can impair accuracy. The paper introduces AAU-net with a hybrid adaptive attention module that selects robust channel- and spatial representations across receptive-field scales. Experiments on three public datasets, together with robustness and external validation analyses, report better segmentation and generalization performance than compared methods.
Problem
Breast-lesion segmentation is challenged by complex ultrasound patterns, similar intensity distributions, variable tumor morphologies, and blurred boundaries.
Method
AAU-net replaces conventional convolution with a hybrid adaptive attention module combining multiple kernel sizes with channel and spatial self-attention.
Results
Experiments on three public breast ultrasound datasets report consistent segmentation improvements over strong baseline and state-of-the-art segmentation methods, with robustness and external validation supporting generalization.
Takeaways & Limitations
The hybrid adaptive attention module is reported to select robust representations across channel and spatial dimensions and can be flexibly applied to existing network frameworks.
Abstract
from arXiv · showhide
Various deep learning methods have been proposed to segment breast lesion from ultrasound images. However, similar intensity distributions, variable tumor morphology and blurred boundaries present challenges for breast lesions segmentation, especially for malignant tumors with irregular shapes. Considering the complexity of ultrasound images, we develop an adaptive attention U-net (AAU-net) to segment breast lesions automatically and stably from ultrasound images. Specifically, we introduce a hybrid adaptive attention module, which mainly consists of a channel self-attention block and a spatial self-attention block, to replace the traditional convolution operation. Compared with the conventional convolution operation, the design of the hybrid adaptive attention module can help us capture more features under different receptive fields. Different from existing attention mechanisms, the hybrid adaptive attention module can guide the network to adaptively select more robust representation in channel and space dimensions to cope with more complex breast lesions segmentation. Extensive experiments with several state-of-the-art deep learning segmentation methods on three public breast ultrasound datasets show that our method has better performance on breast lesion segmentation. Furthermore, robustness analysis and external experiments demonstrate that our proposed AAU-net has better generalization performance on the segmentation of breast lesions. Moreover, the hybrid adaptive attention module can be flexibly applied to existing network frameworks.
I. INTRODUCTION
Breast ultrasound lesion segmentation is challenged by complex image patterns, variable tumor morphology, blurred boundaries, and similar surrounding tissue. The paper addresses these limitations with AAU-net and a hybrid adaptive attention module designed to select robust representations across channel and spatial dimensions.
- Segmentation challenges: Complex ultrasound patterns, variable tumor morphologies, blurred boundaries, and similar surrounding tissue interfere with breast-lesion segmentation accuracy.These challenges are especially problematic for small and malignant tumors, and some lesions may even be missed.
- Prior approaches: Existing approaches commonly enlarge receptive fields or use attention mechanisms to refine breast-lesion segmentation.Dilated convolutions and attention-based U-net variants are cited as representative strategies.
- Prior approaches: Dilated convolutions alone cannot fully address perturbations from surrounding tissues and blurred boundaries, while existing attention methods may remain tied to fixed receptive fields.Prior combined approaches still have fixed receptive-field sizes and single-attention operations.
- Proposed method: The hybrid adaptive module aggregates multiple kernels and uses channel and spatial self-attention to adaptively select receptive fields at different scales.This design is intended to provide more robust representations than conventional fixed-receptive-field convolutions.
- Proposed method: AAU-net applies the hybrid adaptive module to breast-lesion segmentation and learns generic representations from breast ultrasound images.The network is presented as a novel adaptive attention U-net for this task.
- Experimental evidence: Experiments on three public breast ultrasound datasets report consistent segmentation improvements over a strong baseline and state-of-the-art medical-image segmentation methods.The reported evaluation is described as extensive, but no quantitative values are supplied in the selected passages.
II. METHODOLOGY
AAU-net retains U-net’s four-level encoder–decoder structure but replaces conventional convolution layers with hybrid adaptive attention modules. These modules combine multiple receptive-field scales with channel and spatial constraints to improve adaptation to varied ultrasound lesions.
- AAU-net architecture: AAU-net preserves U-net’s U-shaped architecture with four down-sampling, four up-sampling, and skip-connection operations.Each encoding and decoding stage contains two HAAMs.
- HAAM design: The hybrid adaptive attention module replaces the original convolution layer and uses parallel 3 3, 5 5, and dilated 3 3 convolutions.The dilated convolution uses a dilation rate of 3.
- Adaptive representation: HAAM integrates channel and spatial dimension constraints to learn more robust representations from breast ultrasound images.The channel and spatial self-attention blocks are described in subsequent methodology sections.
- Multi-scale receptive fields: Different convolutional kernel sizes provide receptive fields at different scales, improving the network’s adaptability to different input images.The cited figure compares receptive fields from 5 5 convolution, stacked 3 3 convolutions, and dilated convolution.
B. Channel Self-Attention Block
The channel self-attention block adaptively selects representative features from multiple receptive fields. It generates complementary channel attention maps and uses them to calibrate and integrate feature maps for the next stage.
- Purpose: The channel self-attention block guides the network to select representative features along the channel dimension.Its input is the combined feature maps produced from different receptive fields.
- Feature compression: Global average pooling compresses the combined feature maps into a 1 1 feature representation before fully connected processing.The combined 5 5 and dilated-convolution feature maps are integrated by element-wise addition before pooling.
- Attention-map generation: A fully connected layer, batch normalization, ReLU, and sigmoid activation produce the channel attention map.The fully connected weight matrix is denoted by fc W.
- Feature calibration: The attention map and its complement assign opposing importance weights to channels from different receptive fields.The complementary map is defined as 1 − α, and the calibrated feature maps are integrated as input to the next stage.
C. Spatial Self-Attention Block
The spatial self-attention block refines objective-location information by producing spatial attention maps from fused channel-attended and convolutional features. Complementary spatial maps calibrate the features before convolution produces the block output.
- Inputs and fusion: The spatial self-attention block uses features from the 3 3 convolution layer and channel self-attention block as input.These inputs are fused before spatial attention-map generation.
- Attention-map generation: ReLU, 1 1 convolution, and sigmoid operations generate a spatial attention map from the fused feature maps.The resulting map represents spatial information used to refine objective locations.
- Spatial calibration: The spatial attention map and its complement encode the importance of spatial information at corresponding voxels.The complement is defined as 1 − β, and the maps are resampled to match the feature-map channel dimensions.
- Output: Calibrated feature maps are concatenated and passed through a convolution operation to produce the hybrid adaptive attention output.The output is also the output of the entire HAAM.
D. Loss Function
The network uses binary cross entropy as its training loss for the two-class breast-lesion segmentation task. BCE measures the direct difference between predicted masks and ground-truth pixel labels.
- BCE definition: Binary cross entropy is a widely used loss function for two-class image segmentation.It reflects the direct difference between predicted masks and ground-truth labels.
- Inputs: BCE uses ground-truth pixel labels and predicted mask values to compute the segmentation loss.Both labels and predictions are represented as values in [0,1].
- Training objective: The study uses BCE loss to train the segmentation network.The loss is applied to the network’s breast-lesion mask predictions.
A. BUS Datasets
The study evaluates segmentation networks on three public BUS datasets with different scales, using BUSI and Dataset B for experiments and STU for external validation.
- BUSI contains 780 images acquired at Baheya Hospital using two ultrasound systems, with an average size of 500 × 500 pixels.
- Dataset B contains 163 images collected with a Siemens ACUSON Sequoia C512 system, averaging 760 × 570 pixels.
- Evaluation reports Jaccard, Precision, Recall, Specificity, Dice, AUC, Hausdorff distance, average boundary distance, and average symmetric surface distance.
- STU contains 42 images acquired with a GE Voluson E10 system and is used only for external validation because of its small sample size.
B. Experimental Settings
Experiments combine cross-validation, ablation studies, comparisons with competing methods, and robustness analysis, while restricting boundary metrics to external validation for fair evaluation.
- The study uses three datasets for ablation, state-of-the-art comparisons, and robustness analysis, with four-fold cross-validation on BUSI and Dataset B.
- Training uses Adam with a 0.001 initial learning rate, while the selected epoch and batch sizes are 50 and 12.
- The evaluation includes nine segmentation metrics, but HD, ABD, and ASSD are used only in external validation because failed detections undermine boundary-based fairness.
- Architecture Ablation: Architecture ablation treats U-net as the benchmark and reports component effects on BUSI and Dataset B.
- Parameter Ablation: Parameter ablation finds that both smaller and larger receptive fields are not beneficial, supporting the selected convolution-kernel and dilation settings.
B. Comparison with State-of-the-Art Methods
AAU-net is compared with established segmentation methods across standard, lesion-type, external-validation, and normal-image settings. The reported results consistently favor its segmentation confidence, robustness, and generalization.
- BUSI and Dataset B: AAU-net achieves the highest AUC values on both BUSI and Dataset B compared with the other segmentation methods.
- BUSI and Dataset B: Visual results show that AAU-net produces masks closer to ground truth while alleviating effects from tumor size, surrounding tissue, cascades, and heterostructure.
- Benign and Malignant Lesions: AAU-net achieves higher scores on both malignant and benign BUSI lesions, with competitive performance on malignant-lesion segmentation.
- External Validation: External validation uses models trained on Dataset B and tests them on STU to assess generalization across sites and datasets.
- BUSI with Normal Images: Adding normal BUSI images severely affects segmentation-network performance compared with BUSI without normal images.
- Robustness: Compared methods show inconsistent performance across BUSI and Dataset B, whereas AAU-net is reported to have better robustness and generalization ability.
V. DISCUSSION
The discussion attributes AAU-net’s advantages to adaptive attention and broader feature handling, while acknowledging persistent errors on complex ultrasound images and boundaries.
- Ablation results indicate that the selected AAU-net components and parameters achieve the best breast-lesion segmentation performance.
- U-net variants benefit from skip connections, while AAU-net further improves performance through a larger receptive field and spatial self-attention block.
- Visual analysis identifies missed or false detections for small tumors, similar-intensity tissue, heterogeneous lesions, and blurred or cascaded images.
- Robustness experiments report consistently good AAU-net results across BUSI and Dataset B and highlight the hybrid adaptive attention module’s generalization advantage.
- The authors plan boundary constraints and data augmentation to reduce remaining detection errors and improve contour accuracy.
VI. CONCLUSION
The paper introduces HAAM within AAU-net to adaptively select robust channel- and space-dimensional representations for complex breast-lesion segmentation. Comparative, robustness, and external-validation experiments report better segmentation performance than several state-of-the-art methods.
- AAU-net incorporates a novel hybrid adaptive attention module for breast-lesion segmentation.HAAM is used to construct the adaptive attention U-net.
- HAAM adaptively selects robust representations in both channel and space dimensions for complex lesions.
- Comparative experiments with several state-of-the-art segmentation methods report better breast-lesion segmentation performance for AAU-net.
- Robustness analysis and external validation demonstrate better generalization performance for breast-lesion segmentation.