Introduction
Accurate segmentation of breast lesions in ultrasound (US) videos is a critical component in the diagnostic pipeline for breast cancer, particularly in assessing axillary lymph node (ALN) metastasis. Despite its clinical importance, the development of robust automated segmentation models has been hindered by two primary obstacles: data scarcity and the inherent technical challenges of ultrasound imaging. Existing datasets are frequently small or private, leading to models that struggle to generalize. Furthermore, ultrasound images are characterized by speckle noise, low contrast, and blurry boundaries, making it difficult for standard computer vision architectures to maintain temporal consistency as lesions move or deform across video frames.
To address these challenges, this research introduces a significant new benchmark dataset and a novel architecture: the Frequency and Localization Feature Aggregation Network (FLA-Net). By leveraging frequency-domain processing and localization-based contrastive learning, FLA-Net establishes a new state-of-the-art for breast lesion segmentation.
The FLA-Net Architecture
FLA-Net is designed to overcome the limitations of purely spatial-domain models through three key architectural innovations:
1. Frequency-based Feature Aggregation (FFA)
Traditional convolutional neural networks (CNNs) operate in the spatial domain, often requiring deep stacks of layers to achieve a global receptive field. FLA-Net utilizes a Fast Fourier Transform (FFT) to shift features into the spectral domain. Because a single modification in the frequency domain influences the entire spatial image, this module provides the model with an inherent global context. This allows the network to effectively "see through" local speckle noise and better integrate information from adjacent frames.
2. Two-branch Decoder
Rather than producing a segmentation mask in isolation, FLA-Net employs a dual-branch decoder strategy. One branch is dedicated to predicting a localization heatmap (identifying the "where"), while the other focuses on the precise segmentation mask (identifying the "what"). The features from the localization branch are fed back into the segmentation branch, serving as a spatial prior that guides the model in regions where lesion boundaries may be weak or ambiguous.
3. Location-based Contrastive Loss
To ensure temporal stability, the researchers implemented a variant of triplet loss focused on lesion coordinates. This loss function encourages the model to minimize the distance between predicted lesion locations in neighboring frames of the same video (attraction) while maximizing the distance between locations in frames from different patients (repulsion). This approach specifically targets the physical consistency of the lesion’s trajectory throughout a video sequence.
A New Benchmark for Ultrasound Analysis
A major contribution of this work is the release of a large-scale ultrasound video dataset. Comprising 572 videos and 34,300 annotated frames, it represents a substantial increase in scale over previous public benchmarks. The dataset includes comprehensive annotations for:
- Bounding boxes and segmentation masks.
- Lesion classification (benign vs. malignant).
- Axillary lymph node (ALN) metastasis status.
To evaluate the generalizability of FLA-Net, the authors also benchmarked the model on two public video polyp segmentation datasets: CVC-300-TV and CVC-612-V.
Experimental Results
FLA-Net was compared against nine state-of-the-art image-based and video-based segmentation methods. The results demonstrate superior performance across multiple metrics:
- Breast Lesion Segmentation: On the new US video dataset, FLA-Net achieved a Dice score of 0.789, significantly outperforming the previous best-performing model, DCFNet (0.762). It also recorded the lowest Mean Absolute Error (MAE) at 0.033.
- Generalization (Polyps): On the CVC-300-TV dataset, the model improved the Dice score from 0.840 to 0.874 and the Enhanced-alignment measure ($E_\phi$) from 0.921 to 0.969.
- Efficiency: Despite the integration of frequency-domain modules, FLA-Net maintains lower time and space complexity compared to several existing heavy-duty video architectures, making it a more viable candidate for clinical integration.
Discussion and Future Work
While FLA-Net represents a significant advancement, the study identifies areas for further development. Ultrasound imaging is highly operator-dependent; the impact of variations in probe pressure and angle on model performance remains an area for deeper investigation. Additionally, while the model is computationally efficient, future work should establish specific frames-per-second (FPS) benchmarks on standard clinical hardware to ensure real-time viability.
The success of the frequency-domain approach in this context suggests broad potential for other medical imaging tasks. Future research may extend these techniques to echocardiography or surgical video analysis, where motion blur and noise are similarly persistent challenges.
Summary updated