Abstract:Breast UltraSound (BUS) imaging is a commonly used diagnostic tool in the field of counter fighting breast diseases, especially for early detection and diagnosis of breast cancer. Due to the inherent characteristics of ultrasound images such as blurry boundaries and diverse tumor morphologies, it is challenging for doctors to manually segment breast tumors. In recent years, the Convolutional Neural Network (CNN) technology has been widely applied to automatically segment BUS images. However, due to the inherent limitations of CNNs in capturing global contextual information, it is difficult to capture the full context. To address this issue, the paper proposes a novel BGRD-TransUNet model for breast lesion segmentation, based on TransUNet. The proposed model, first, replaces the original ResNet50 backbone network of TransUNet with DenseNet121 for initial feature extraction. Next, newly designed Residual Multi-Scale Feature Modules (RMSFMs) are employed to extract features from various layers of DenseNet121, thus capturing richer features within specific layers. Thirdly, a Boundary Guidance (BG) network is added to enhance the contour information of BUS images. Additionally, newly designed Boundary Attentional Feature Fusion Modules (BAFFMs) are used to integrate edge information and features extracted through RMSFMs. Finally, newly designed Parallel Channel and Spatial Attention Modules (PCSAMs) are used to refine feature extraction using channel and spatial attention. An extensive experimental testing performed on two public datasets demonstrates that the proposed BGRD-TransUNet model outperforms all state-of-the-art medical image segmentation models, participating in the experiments, according to all evaluation metrics used (except for few separate cases), including the two most important and widely used metrics in the field of medical image segmentation, namely the Intersection over Union (IoU) and Dice Similarity Coefficient (DSC). More specifically, on the BUSI dataset and dataset B, BGRD-TransUNet achieves IoU values of 76.77% and 86.61%, and DSC values of 85.08% and 92.47%, respectively, which are higher by 7.27 and 3.64, and 5.81 and 2.54 percentage points, than the corresponding values achieved by the baseline (TransUNet).

What problem does this paper attempt to address?

The paper attempts to address the challenges faced in tumor segmentation in breast ultrasound (BUS) images. Specifically, due to the inherent characteristics of ultrasound images, such as blurred boundaries and diverse tumor morphologies, it is very difficult for doctors to manually segment breast tumors. Although convolutional neural network (CNN) technology has been widely applied to automatically segment BUS images, it is difficult to comprehensively capture all relevant information due to the limitations of CNN in capturing global contextual information. To overcome these issues, the paper proposes a new model based on TransUNet—BGRD-TransUNet, specifically for breast lesion segmentation. ### Main Issues: 1. **Challenges of Manual Segmentation**: Due to the blurred boundaries and diverse tumor morphologies in ultrasound images, it is difficult for doctors to manually segment breast tumors. 2. **Limitations of CNN**: Traditional CNNs have limitations in capturing global contextual information, making it difficult to comprehensively capture all relevant information. 3. **Small Dataset**: Medical image datasets (especially BUS datasets) are usually small in scale, which can easily lead to model overfitting. 4. **Loss of Edge Information**: During the segmentation process, spatial information is easily lost due to multiple upsamplings. This problem is particularly severe in ultrasound images where the boundaries are already quite blurred. ### Solutions: - **Model Improvements**: The paper proposes the BGRD-TransUNet model, which addresses the above issues through the following improvements: 1. **Replacing the Backbone Network**: The original ResNet50 backbone network in TransUNet is replaced with DenseNet121, utilizing DenseNet121's parameter sharing and feature reuse characteristics to reduce the risk of overfitting. 2. **Residual Multi-Scale Feature Modules (RMSFMs)**: New RMSFMs are designed to expand the receptive field and capture richer feature information. 3. **Deformable Atrous Spatial Pyramid Pooling Module (DASPPM)**: A new DASPPM module is introduced to enhance the ability to extract complex shapes. 4. **Boundary Guidance Network (BG)**: A new boundary guidance network is designed to specifically train boundary features, addressing the issue of blurred boundaries in ultrasound images. 5. **Boundary Attention Feature Fusion Modules (BAFFMs)**: New BAFFMs are introduced to achieve attention-driven feature fusion through a multi-scale channel attention mechanism, alleviating the semantic differences between shallow encoders and decoders. 6. **Parallel Channel and Spatial Attention Modules (PCSAMs)**: New PCSAMs are used in the decoder to simultaneously consider channel information and spatial information, better capturing key features in the image. Through these improvements, experimental results of the BGRD-TransUNet model on two public datasets show that it outperforms existing state-of-the-art medical image segmentation models on all evaluation metrics, particularly excelling in the two most important metrics: Intersection over Union (IoU) and Dice Similarity Coefficient (DSC).

BGRD-TransUNet: A Novel TransUNet-Based Model for Ultrasound Breast Lesion Segmentation

An RDAU-NET model for lesion segmentation in breast ultrasound images

Boundary-guided and Region-aware Network with Global Scale-adaptive for Accurate Segmentation of Breast Tumors in Ultrasound Images

A deep supervised transformer U‐shaped full‐resolution residual network for the segmentation of breast ultrasound image

Rethinking the unpretentious U-net for medical ultrasound image segmentation

Automatic tumor segmentation in breast ultrasound images using a dilated fully convolutional network combined with an active contour model.

Fully automatic tumor segmentation of breast ultrasound images with deep learning

Attention gate and dilation U-shaped network (GDUNet): an efficient breast ultrasound image segmentation network with multiscale information extraction

CBAM-RIUnet: Breast Tumor Segmentation With Enhanced Breast Ultrasound and Test-Time Augmentation

GBE-Net: Global Boundary Enhancement Network for breast lesion segmentation in ultrasound images

BiFTransNet: A unified and simultaneous segmentation network for gastrointestinal images of CT & MRI

EG-TransUNet: a transformer-based U-Net with enhanced and guided models for biomedical image segmentation

Attention based UNet model for breast cancer segmentation using BUSI dataset

DDRA-Net: Dual-Channel Deep Residual Attention UPerNet for Breast Lesions Segmentation in Ultrasound Images

Multi-task learning for segmentation and classification of breast tumors from ultrasound images

BUSnet: A Deep Learning Model of Breast Tumor Lesion Detection for Ultrasound Images

TBUnet: A Pure Convolutional U-Net Capable of Multifaceted Feature Extraction for Medical Image Segmentation

Automatic Boundary Detection in Breast Ultrasound Images Based on Improved Pulse Coupled Neural Network and Active Contour Model

BGM-Net: Boundary-Guided Multiscale Network for Breast Lesion Segmentation in Ultrasound

WATUNet: A Deep Neural Network for Segmentation of Volumetric Sweep Imaging Ultrasound