Source-linked AI summary
Global Guidance Network for Breast Lesion Segmentation in Ultrasound Images
Cheng Xue, Lei Zhu, Huazhu Fu, Xiaowei Hu, Xiaomeng Li, Hai Zhang, Pheng Ann Heng
TL;DR
Breast ultrasound lesion segmentation is challenged by noisy images, unclear boundaries, inhomogeneous lesion regions, and CNNs’ limited long-range context. The paper proposes GG-Net, combining globally guided spatial and channel non-local features with boundary detection. It reports more accurate breast lesion segmentation than competing methods and satisfactory prostate segmentation performance.
Problem
Breast ultrasound segmentation is difficult because lesion appearance and boundaries are ambiguous, while local CNN operations inadequately capture long-range dependencies.
Method
GG-Net integrates multi-layer CNN features to guide non-local learning in spatial and channel domains and predicts additional lesion boundary maps for refinement.
Results
GG-Net more accurately segments breast lesions than competing state-of-the-art methods on a public and a collected dataset, with satisfactory performance on ultrasound prostate segmentation.
Takeaways & Limitations
The reported results support using global guidance and boundary information to improve ultrasound lesion segmentation across breast datasets and in prostate application.
Takeaways & Limitations
The network can fail on very large lesions with complicated internal intensity distributions or unclear boundaries, and the authors identify elastography integration as future work.
Abstract
from arXiv · showhide
Automatic breast lesion segmentation in ultrasound helps to diagnose breast cancer, which is one of the dreadful diseases that affect women globally. Segmenting breast regions accurately from ultrasound image is a challenging task due to the inherent speckle artifacts, blurry breast lesion boundaries, and inhomogeneous intensity distributions inside the breast lesion regions. Recently, convolutional neural networks (CNNs) have demonstrated remarkable results in medical image segmentation tasks. However, the convolutional operations in a CNN often focus on local regions, which suffer from limited capabilities in capturing long-range dependencies of the input ultrasound image, resulting in degraded breast lesion segmentation accuracy. In this paper, we develop a deep convolutional neural network equipped with a global guidance block (GGB) and breast lesion boundary detection (BD) modules for boosting the breast ultrasound lesion segmentation. The GGB utilizes the multi-layer integrated feature map as a guidance information to learn the long-range non-local dependencies from both spatial and channel domains. The BD modules learn additional breast lesion boundary map to enhance the boundary quality of a segmentation result refinement. Experimental results on a public dataset and a collected dataset show that our network outperforms other medical image segmentation methods and the recent semantic segmentation methods on breast ultrasound lesion segmentation. Moreover, we also show the application of our network on the ultrasound prostate segmentation, in which our method better identifies prostate regions than state-of-the-art networks.
1. Introduction
Breast ultrasound lesion segmentation is difficult because of noise, ambiguous boundaries, inhomogeneous lesion intensities, irregular shapes, and limited long-range modeling in conventional CNNs. GG-Net addresses these issues with globally guided spatial and channel features plus boundary detection, outperforming state-of-the-art methods on breast and prostate ultrasound segmentation.
- Motivation: Speckle noise, strong shadows, inhomogeneous lesion intensities, ambiguous boundaries, and irregular shapes make automatic breast lesion segmentation challenging.These difficulties are illustrated by inhomogeneous lesion interiors, similar lesion and background appearances, and irregular lesion shapes.
- Motivation: Conventional CNNs primarily use local operations, limiting their ability to capture long-range dependencies and global contextual information in ultrasound images.Distant pixels can resemble lesions, while existing receptive-field expansion and feature-fusion methods do not fully capture global context.
- Method: GG-Net integrates shallow and deep CNN features into multi-layer integrated features that guide non-local learning across spatial and channel domains.The global guidance block uses these integrated features to complement lesion details that deep layers may neglect.
- Method: The breast lesion boundary detection module predicts additional boundary maps in shallow CNN layers to refine segmentation boundaries.The module is designed to preserve boundary details that may be missing from deeper representations.
- Results: GG-Net outperforms state-of-the-art medical image segmentation methods on two breast ultrasound lesion datasets.The reported comparison covers both a public dataset and a collected dataset.
- Results: GG-Net also achieves satisfactory performance when applied to ultrasound prostate segmentation.This extends the reported evaluation beyond breast lesion segmentation.
2. Related works
Breast ultrasound lesion segmentation has progressed from hand-crafted and conventional learning methods to CNN-based approaches that incorporate broader contextual information. GG-Net combines multi-layer contextual guidance, spatial- and channel-wise dependencies, and boundary learning to address limitations of local features and improve segmentation quality.
- Breast lesion segmentation is challenging because ultrasound contains speckle artifacts, low contrast, shadows, blurry boundaries, and varied lesion shapes.
- Traditional region-, deformable-model-, graph-, and learning-based methods commonly represent local intensity variation with texture features.
- CNNs have achieved strong performance in medical image analysis and have been applied to breast mass detection, density estimation, and ultrasound segmentation.
- ASPP, pyramid pooling, multi-scale feature fusion, deformable convolutions, attention, and non-local operations were developed to enlarge receptive fields and capture contextual information.
- GG-Net unifies multi-layer contextual features with spatial-wise and channel-wise long-range dependencies while embedding information from different CNN layers.
- The network also uses boundary detection to learn additional lesion boundary maps, producing segmentation results with higher-quality boundaries.
3. Methodology
GG-Net processes breast ultrasound images end to end, combining multi-scale CNN features and ASPP with global guidance to refine features using long-range dependencies. Its final output is a segmented lesion mask.
- GG-Net takes a breast ultrasound image as input and produces a segmented mask in an end-to-end manner.
- A CNN generates multi-level feature maps at different spatial resolutions, while ASPP enhances their receptive field.
- The GGB refines features by learning long-range dependencies under guidance from integrated information across CNN layers.
3.1. Global Guidance Block
The global guidance block integrates shallow and deep features to learn long-range dependencies across spatial positions and feature channels. Its spatial pathway combines input-feature and guidance-map similarities to produce refined features.
- Local convolutional operations can miss long-range context, while ultrasound speckles and shadows may resemble lesions when receptive fields are limited.
- Shallow CNN layers retain lesion details but include non-lesion regions, whereas deep layers suppress non-lesion regions but can lose local details.
- The method resizes feature maps from the first four CNN layers and concatenates them into a multi-layer integrated feature map used as guidance.
- The spatial-wise GGB computes position similarities from the ASPP output and from the guidance map, then combines them into a guided similarity matrix.
- The guided similarity matrix is applied to transformed input features and added back to the input to generate the refined output feature map.
3.2. Loss Function
The channel-wise GGB emphasizes informative guidance channels before learning inter-channel dependencies and refining the input features.
- Two fully connected layers and sigmoid activation transform β into a coefficient vector Vλ.
- The coefficient vector weights guidance channels, after which channel similarity maps guide transformation of the input features.
3.3. Breast Lesion Boundary Detection Module
The boundary detection module adds multi-level boundary supervision to improve breast lesion segmentation and boundary quality alongside the final GGB-based prediction.
- The BD module addresses inaccurate lesion boundaries by predicting additional boundary maps and adding a boundary prediction loss.
- Shallow CNN layers jointly locate lesions and detect boundaries, producing four boundary maps and four intermediate segmentation results.
- The training objective combines segmentation losses at four CNN layers, boundary losses, and the final segmentation loss.
- The weights λ1 = 1 and λ2 = 10 balance intermediate segmentation, boundary, and final segmentation losses.
- The boundary loss is computed as mean squared error between predicted boundary maps D_i and the ground-truth boundary map B_G.
3.4. Implementation
Training initializes the feature extractor from ImageNet-pretrained ResNeXt and optimizes the full network with SGD; inference applies CRF refinement.
- The feature extractor is initialized with pretrained ResNeXt on ImageNet, while other parameters use random initialization.
- SGD training uses momentum 0.9, weight decay 0.0001, mini-batches of 4, and 100 epochs.
- The initial learning rate is 0.001 and is multiplied by 0.1 after every 50 epochs.
- Testing outputs the refined dual-guided non-local segmentation result after conditional random field processing.
- The network has 55M trainable parameters and an inference time of 0.039 seconds per image.
4. Experiments
Experiments evaluate GG-Net on two breast ultrasound datasets using ablation studies and comparisons with segmentation methods. The results show benefits from global guidance, boundary detection, and long-range dependency modeling.
- Experimental setup: Experiments use the public BUSI dataset, a collected dataset, seven segmentation metrics, and radiologist-derived ground truths.BUSI contains 780 images from 600 patients, while the collected dataset contains 632 images from 200 patients.
- Ablation analysis: The ablation baseline is DeeplabV3+ with a ResNeXt backbone, constructed without GGB or the boundary-detection module.The ablation study evaluates spatial-wise GGB, channel-wise GGB, and BD components.
- Ablation analysis: Learning long-range dependencies improves breast lesion segmentation compared with the baseline.This comparison is reported for the spatial and channel non-local components in Table 1.
- Ablation analysis: Adding MLIF guidance improves both spatial and channel non-local blocks, while combining spatial-wise and channel-wise guidance outperforms either alone.The guidance information helps capture long-range position dependencies for breast lesion segmentation.
- Ablation analysis: The full method achieves the best ablation accuracy, and adding BD further improves visual segmentation by generating refined boundaries.Visual comparisons show improvements from basic to basic+GGB and from basic+GGB to basic+GGB+BD.
- Quantitative comparisons: GG-Net outperforms compared segmentation methods across the reported metrics on the collected and BUSI datasets.It has larger Jaccard, Dice, Accuracy, Recall, and Precision values and smaller HD and ABD values than competitors.
- Dataset and visual comparisons: Performance is better on the collected dataset than BUSI across all seven metrics, which the authors attribute to higher ultrasound image quality.The BUSI experiment also includes normal cases in the training data, and visual comparisons report fewer omitted details or false lesion regions for GG-Net.
5. Application
The network is additionally evaluated for ultrasound prostate segmentation under the DAF experimental setting. It outperforms the compared state-of-the-art networks on almost all reported metrics.
- Prostate segmentation: The network is retrained and tested for ultrasound prostate segmentation using the DAF training and testing sets.Comparisons use Jaccard, Dice, Recall, and Precision.
- Prostate segmentation: GG-Net outperforms U-Net, FCN, BCRNN, and DAF on almost all four prostate segmentation metrics.The quantitative results are summarized in Table 7.
6. Discussions
The discussion links GG-Net’s gains to global guidance for non-local feature learning and boundary detection for lesion detail. It also reports statistical support, clinical relevance, and failure cases.
- Discussion: Traditional CNNs use local convolutions with limited receptive fields, motivating non-local dependency modeling for ultrasound lesion segmentation.Prior non-local blocks were embedded only in deep layers, which can lack parts of lesion regions captured by shallow layers.
- Discussion: GG-Net combines spatial and channel guidance with boundary detection, and reports significant improvement over compared methods on two datasets.The paper reports p-value <0.05 in Table 8.
- Clinical relevance: The method performs better on relatively less obvious lesions, which commonly have low contrast, shadows, and blurry boundaries.The authors identify this setting as especially relevant to breast ultrasound clinical practice.
- Future work: Future work will use elastography images to further improve ultrasound breast lesion segmentation.The paper notes that elastography images encode tissue density.
- Limitations: Failure cases occur when lesions are very large, have complicated internal intensity distributions, or have unclear boundaries.The network may falsely label non-lesion regions or omit parts of the lesion.
- Statistical analysis: Across seven metrics, most paired comparisons have p-values below 0.05, although several Accuracy comparisons are closer to the significance threshold.The authors state that the corresponding Accuracy performance is similar for those comparisons.
7. Conclusion
The paper presents GG-Net for ultrasound breast lesion segmentation using global guidance and boundary detection. Experiments report higher segmentation accuracy than state-of-the-art methods on breast and prostate ultrasound tasks.
- Conclusion: GG-Net combines multi-layer context guidance for spatial and channel non-local features with an additional breast lesion boundary map.The network is evaluated on public and collected breast ultrasound datasets and also applied to prostate segmentation.
- Conclusion: Experimental results report more accurate breast lesion segmentation and higher prostate segmentation accuracy than state-of-the-art methods.The conclusion states that GG-Net outperforms all competitors in the reported applications.